Real-Time

Digital Signal Processing
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
Real-Time
Digital Signal Processing
Implementations, Applications, and
Experiments with the TMS320C55X
Sen M Kuo
Northern Illinois University, DeKalb, Illinois, USA
Bob H Lee
Texas Instruments, Inc., Schaumburg, Illinois, USA
JOHN WILEY & SONS, LTD.
Chichester

New York

Weinheim

Brisbane

Singapore

Toronto
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
Copyright #2001 by John Wiley & Sons, Ltd
Baffins Lane, Chichester,
West Sussex, PO 19 1UD, England
National 01243 779777
International (‡44) 1243 779777
e-mail (for orders and customer service enquiries): cs-books@wiley.co.uk
Visit our Home Page on http://www.wiley.co.uk
or
http://www.wiley.com
All Rights Reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in
any form or by any means, electronic, mechanical, photocopying, recording, scanning or otherwise, except under the
terms of the Copyright Designs and Patents Act 1988 or under the terms of a licence issued by the Copyright
Licensing Agency, 90 Tottenham Court Road, London, W1P 9HE, UK, without the permission in writing of the
Publisher, with the exception of any material supplied specifically for the purpose of being entered and executed on a
computer system, for exclusive use by the purchaser of the publication.
Neither the authors nor John Wiley & Sons Ltd accept any responsibility or liability for loss or damage occasioned to
any person or property through using the material, instructions, methods or ideas contained herein, or acting or
refraining from acting as a result of such use. The authors and Publisher expressly disclaim all implied warranties,
including merchantability of fitness for any particular purpose. There will be no duty on the authors or Publisher to
correct any errors or defects in the software.
Designations used by companies to distinguish their products are often claimed as trademarks. In all instances where
John Wiley & Sons is aware of a claim, the product names appear in initial capital or capital letters. Readers,
however, should contact the appropriate companies for more complete information regarding trademarks and
registration.
Other Wiley Editorial Offices
John Wiley & Sons, Inc., 605 Third Avenue,
New York, NY 10158±0012, USA
Wiley-VCH Verlag GmbH
Pappelallee 3, D-69469 Weinheim, Germany
Jacaranda Wiley Ltd, 33 Park Road, Milton,
Queensland 4064, Australia
John Wiley & Sons (Canada) Ltd, 22 Worcester Road
Rexdale, Ontario, M9W 1L1, Canada
John Wiley & Sons (Asia) Pte Ltd, 2 Clementi Loop #02±01,
Jin Xing Distripark, Singapore 129809
Library of Congress Cataloging-in-Publication Data
Kuo, Sen M. (Sen-Maw)
Real-time digital signal processing: implementations, applications, and experiments
with the TMS320C55x / Sen M. Kuo, Bob H. Lee
p. cm.
Includes bibliographical references and index.
ISBN 0±470±84137±0
1. Signal processingÐDigital techniques. 2. Texas Instruments TMS320 series
microprocessors. I. Lee, Bob H. II. Title.
TK5102.9 .K86 2001
621.382
0
2Ðdc21 2001026651
British Library Cataloguing in Publication Data
A catalogue record for this book is available from the British Library
ISBN 0 470 84137 0
Typeset by Kolam Information Services Pvt. Ltd, Pondicherry, India
Printed and bound in Great Britain by Antony Rowe Ltd
This book is printed on acid-free paper responsibly manufactured from sustainable forestry,
in which at least two trees are planted for each one used for paper production.
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
To my wife Paolien, and children Jennifer, Kevin,
and Kathleen.
± Sen M. Kuo
To my dear wife Vikki and daughter Jenni.
± Bob H. Lee
Contents
Preface xv
1 Introduction to Real-Time Digital Signal Processing 1
1.1 Basic Elements of Real-Time DSP Systems 2
1.2 Input and Output Channels 3
1.2.1 Input Signal Conditioning 3
1.2.2 A/D Conversion 4
1.2.3 Sampling 5
1.2.4 Quantizing and Encoding 7
1.2.5 D/A Conversion 9
1.2.6 Input/Output Devices 9
1.3 DSP Hardware 11
1.3.1 DSP Hardware Options 11
1.3.2 Fixed- and Floating-Point Devices 13
1.3.3 Real-Time Constraints 14
1.4 DSP System Design 14
1.4.1 Algorithm Development 14
1.4.2 Selection of DSP Chips 16
1.4.3 Software Development 17
1.4.4 High-Level Software Development Tools 18
1.5 Experiments Using Code Composer Studio 19
1.5.1 Experiment 1A ± Using the CCS and the TMS320C55x Simulator 20
1.5.2 Experiment 1B ± Debugging Program on the CCS 25
1.5.3 Experiment 1C ± File Input and Output 28
1.5.4 Experiment 1D ± Code Efficiency Analysis 29
1.5.5 Experiment 1E ± General Extension Language 32
References 33
Exercises 33
2 Introduction to TMS320C55x Digital Signal Processor 35
2.1 Introduction 35
2.2 TMS320C55x Architecture 36
2.2.1 TMS320C55x Architecture Overview 36
2.2.2 TMS320C55x Buses 39
2.2.3 TMS320C55x Memory Map 40
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
2.3 Software Development Tools 40
2.3.1 C Compiler 42
2.3.2 Assembler 44
2.3.3 Linker 46
2.3.4 Code Composer Studio 48
2.3.5 Assembly Statement Syntax 49
2.4 TMS320C55x Addressing Modes 50
2.4.1 Direct Addressing Mode 52
2.4.2 Indirect Addressing Mode 53
2.4.3 Absolute Addressing Mode 56
2.4.4 Memory-Mapped Register Addressing Mode 56
2.4.5 Register Bits Addressing Mode 57
2.4.6 Circular Addressing Mode 58
2.5 Pipeline and Parallelism 59
2.5.1 TMS320C55x Pipeline 59
2.5.2 Parallel Execution 60
2.6 TMS320C55x Instruction Set 63
2.6.1 Arithmetic Instructions 63
2.6.2 Logic and Bits Manipulation Instructions 64
2.6.3 Move Instruction 65
2.6.4 Program Flow Control Instructions 66
2.7 Mixed C and Assembly Language Programming 68
2.8 Experiments ± Assembly Programming Basics 70
2.8.1 Experiment 2A ± Interfacing C with Assembly Code 71
2.8.2 Experiment 2B ± Addressing Mode Experiments 72
References 75
Exercises 75
3 DSP Fundamentals and Implementation
Considerations 77
3.1 Digital Signals and Systems 77
3.1.1 Elementary Digital Signals 77
3.1.2 Block Diagram Representation of Digital Systems 79
3.1.3 Impulse Response of Digital Systems 83
3.2 Introduction to Digital Filters 83
3.2.1 FIR Filters and Power Estimators 84
3.2.2 Response of Linear Systems 87
3.2.3 IIR Filters 88
3.3 Introduction to Random Variables 90
3.3.1 Review of Probability and Random Variables 90
3.3.2 Operations on Random Variables 92
3.4 Fixed-Point Representation and Arithmetic 95
3.5 Quantization Errors 98
3.5.1 Input Quantization Noise 98
3.5.2 Coefficient Quantization Noise 101
3.5.3 Roundoff Noise 102
3.6 Overflow and Solutions 103
3.6.1 Saturation Arithmetic 103
3.6.2 Overflow Handling 104
3.6.3 Scaling of Signals 105
3.7 Implementation Procedure for Real-Time Applications 107
viii CONTENTS
3.8 Experiments of Fixed-Point Implementations 108
3.8.1 Experiment 3A ± Quantization of Sinusoidal Signals 109
3.8.2 Experiment 3B ± Quantization of Speech Signals 111
3.8.3 Experiment 3C ± Overflow and Saturation Arithmetic 112
3.8.4 Experiment 3D ± Quantization of Coefficients 115
3.8.5 Experiment 3E ± Synthesizing Sine Function 117
References 121
Exercises 122
4 Frequency Analysis 127
4.1 Fourier Series and Transform 127
4.1.1 Fourier Series 127
4.1.2 Fourier Transform 130
4.2 The z-Transforms 133
4.2.1 Definitions and Basic Properties 133
4.2.2 Inverse z-Transform 136
4.3 System Concepts 141
4.3.1 Transfer Functions 141
4.3.2 Digital Filters 143
4.3.3 Poles and Zeros 144
4.3.4 Frequency Responses 148
4.4 Discrete Fourier Transform 152
4.4.1 Discrete-Time Fourier Series and Transform 152
4.4.2 Aliasing and Folding 154
4.4.3 Discrete Fourier Transform 157
4.4.4 Fast Fourier Transform 159
4.5 Applications 160
4.5.1 Design of Simple Notch Filters 160
4.5.2 Analysis of Room Acoustics 162
4.6 Experiments Using the TMS320C55x 165
4.6.1 Experiment 4A ± Twiddle Factor Generation 167
4.6.2 Experiment 4B ± Complex Data Operation 169
4.6.3 Experiment 4C ± Implementation of DFT 171
4.6.4 Experiment 4D ± Experiment Using Assembly Routines 173
References 176
Exercises 176
5 Design and Implementation of FIR Filters 181
5.1 Introduction to Digital Filters 181
5.1.1 Filter Characteristics 182
5.1.2 Filter Types 183
5.1.3 Filter Specifications 185
5.2 FIR Filtering 189
5.2.1 Linear Convolution 189
5.2.2 Some Simple FIR Filters 192
5.2.3 Linear Phase FIR Filters 194
5.2.4 Realization of FIR Filters 198
5.3 Design of FIR Filters 201
5.3.1 Filter Design Procedure 201
5.3.2 Fourier Series Method 202
5.3.3 Gibbs Phenomenon 205
CONTENTS ix
5.3.4 Window Functions 208
5.3.5 Frequency Sampling Method 214
5.4 Design of FIR Filters Using MATLAB 219
5.5 Implementation Considerations 221
5.5.1 Software Implementations 221
5.5.2 Quantization Effects in FIR Filters 223
5.6 Experiments Using the TMS320C55x 225
5.6.1 Experiment 5A ± Implementation of Block FIR Filter 227
5.6.2 Experiment 5B ± Implementation of Symmetric FIR Filter 230
5.6.3 Experiment 5C ± Implementation of FIR Filter Using Dual-MAC 233
References 235
Exercises 236
6 Design and Implementation of IIR Filters 241
6.1 Laplace Transform 241
6.1.1 Introduction to the Laplace Transform 241
6.1.2 Relationships between the Laplace and z-Transforms 245
6.1.3 Mapping Properties 246
6.2 Analog Filters 247
6.2.1 Introduction to Analog Filters 248
6.2.2 Characteristics of Analog Filters 249
6.2.3 Frequency Transforms 253
6.3 Design of IIR Filters 255
6.3.1 Review of IIR Filters 255
6.3.2 Impulse-Invariant Method 256
6.3.3 Bilinear Transform 259
6.3.4 Filter Design Using Bilinear Transform 261
6.4 Realization of IIR Filters 263
6.4.1 Direct Forms 263
6.4.2 Cascade Form 266
6.4.3 Parallel Form 268
6.4.4 Realization Using MATLAB 269
6.5 Design of IIR Filters Using MATLAB 271
6.6 Implementation Considerations 273
6.6.1 Stability 274
6.6.2 Finite-Precision Effects and Solutions 275
6.6.3 Software Implementations 279
6.6.4 Practical Applications 280
6.7 Software Developments and Experiments Using the TMS320C55x 284
6.7.1 Design of IIR Filter 285
6.7.2 Experiment 6A ± Floating-Point C Implementation 286
6.7.3 Experiment 6B ± Fixed-Point C Implementation Using Intrinsics 289
6.7.4 Experiment 6C ± Fixed-Point C Programming Considerations 292
6.7.5 Experiment 6D ± Assembly Language Implementations 295
References 297
Exercises 297
7 Fast Fourier Transform and Its Applications 303
7.1 Discrete Fourier Transform 303
7.1.1 Definitions 304
7.1.2 Important Properties of DFT 308
7.1.3 Circular Convolution 311
x CONTENTS
7.2 Fast Fourier Transforms 314
7.2.1 Decimation-in-Time 315
7.2.2 Decimation-in-Frequency 319
7.2.3 Inverse Fast Fourier Transform 320
7.2.4 MATLAB Implementations 321
7.3 Applications 322
7.3.1 Spectrum Estimation and Analysis 322
7.3.2 Spectral Leakage and Resolution 324
7.3.3 Power Density Spectrum 328
7.3.4 Fast Convolution 330
7.3.5 Spectrogram 332
7.4 Implementation Considerations 333
7.4.1 Computational Issues 334
7.4.2 Finite-Precision Effects 334
7.5 Experiments Using the TMS320C55x 336
7.5.1 Experiment 7A ± Radix-2 Complex FFT 336
7.5.2 Experiment 7B ± Radix-2 Complex FFT Using Assembly Language 341
7.5.3 Experiment 7C ± FFT and IFFT 344
7.5.4 Experiment 7D ± Fast Convolution 344
References 346
Exercises 347
8 Adaptive Filtering 351
8.1 Introduction to Random Processes 351
8.1.1 Correlation Functions 352
8.1.2 Frequency-Domain Representations 356
8.2 Adaptive Filters 359
8.2.1 Introduction to Adaptive Filtering 359
8.2.2 Performance Function 361
8.2.3 Method of Steepest Descent 365
8.2.4 The LMS Algorithm 366
8.3 Performance Analysis 367
8.3.1 Stability Constraint 367
8.3.2 Convergence Speed 368
8.3.3 Excess Mean-Square Error 369
8.4 Modified LMS Algorithms 370
8.4.1 Normalized LMS Algorithm 370
8.4.2 Leaky LMS Algorithm 371
8.5 Applications 372
8.5.1 Adaptive System Identification 372
8.5.2 Adaptive Linear Prediction 373
8.5.3 Adaptive Noise Cancellation 375
8.5.4 Adaptive Notch Filters 377
8.5.5 Adaptive Channel Equalization 379
8.6 Implementation Considerations 381
8.6.1 Computational Issues 381
8.6.2 Finite-Precision Effects 382
8.7 Experiments Using the TMS320C55x 385
8.7.1 Experiment 8A ± Adaptive System Identification 385
8.7.2 Experiment 8B ± Adaptive Predictor Using the Leaky LMS Algorithm 390
References 396
Exercises 396
CONTENTS xi
9 Practical DSP Applications in Communications 399
9.1 Sinewave Generators and Applications 399
9.1.1 Lookup-Table Method 400
9.1.2 Linear Chirp Signal 402
9.1.3 DTMF Tone Generator 403
9.2 Noise Generators and Applications 404
9.2.1 Linear Congruential Sequence Generator 404
9.2.2 Pseudo-Random Binary Sequence Generator 406
9.2.3 Comfort Noise in Communication Systems 408
9.2.4 Off-Line System Modeling 409
9.3 DTMF Tone Detection 410
9.3.1 Specifications 410
9.3.2 Goertzel Algorithm 411
9.3.3 Implementation Considerations 414
9.4 Adaptive Echo Cancellation 417
9.4.1 Line Echoes 417
9.4.2 Adaptive Echo Canceler 418
9.4.3 Practical Considerations 422
9.4.4 Double-Talk Effects and Solutions 423
9.4.5 Residual Echo Suppressor 425
9.5 Acoustic Echo Cancellation 426
9.5.1 Introduction 426
9.5.2 Acoustic Echo Canceler 427
9.5.3 Implementation Considerations 428
9.6 Speech Enhancement Techniques 429
9.6.1 Noise Reduction Techniques 429
9.6.2 Spectral Subtraction Techniques 431
9.6.3 Implementation Considerations 433
9.7 Projects Using the TMS320C55x 435
9.7.1 Project Suggestions 435
9.7.2 A Project Example ± Wireless Application 437
References 442
Appendix A Some Useful Formulas 445
A.1 Trigonometric Identities 445
A.2 Geometric Series 446
A.3 Complex Variables 447
A.4 Impulse Functions 449
A.5 Vector Concepts 449
A.6 Units of Power 450
Reference 451
Appendix B Introduction of MATLAB for DSP
Applications 453
B.1 Elementary Operations 453
B.1.1 Initializing Variables and Vectors 453
B.1.2 Graphics 455
B.1.3 Basic Operators 457
B.1.4 Files 459
B.2 Generation and Processing of Digital Signals 460
B.3 DSP Applications 463
B.4 User-Written Functions 465
xii CONTENTS
B.5 Summary of Useful MATLAB Functions 466
References 467
Appendix C Introduction of C Programming for DSP
Applications 469
C.1 A Simple C Program 470
C.1.1 Variables and Assignment Operators 472
C.1.2 Numeric Data Types and Conversion 473
C.1.3 Arrays 474
C.2 Arithmetic and Bitwise Operators 475
C.2.1 Arithmetic Operators 475
C.2.2 Bitwise Operators 476
C.3 An FIR Filter Program 476
C.3.1 Command-Line Arguments 477
C.3.2 Pointers 477
C.3.3 C Functions 478
C.3.4 Files and I/O Operations 480
C.4 Control Structures and Loops 481
C.4.1 Control Structures 481
C.4.2 Logical Operators 483
C.4.3 Loops 484
C.5 Data Types Used by the TMS320C55x 485
References 486
Appendix D About the Software 487
Index 489
CONTENTS xiii
Preface
Real-time digital signal processing (DSP) using general-purpose DSP processors is very
challenging work in today's engineering fields. It promises an effective way to design,
experiment, and implement a variety of signal processing algorithms for real-world
applications. With DSP penetrating into various applications, the demand for high-
performance digital signal processors has expanded rapidly in recent years. Many
industrial companies are currently engaged in real-time DSP research and development.
It becomes increasingly important for today's students and practicing engineers to
master not only the theory of DSP, but equally important, the skill of real-time DSP
system design and implementation techniques.
This book offers readers a hands-on approach to understanding real-time DSP
principles, system design and implementation considerations, real-world applications,
as well as many DSP experiments using MATLAB, C/C++, and the TMS320C55x. This
is a practical book about DSP and using digital signal processors for DSP applications.
This book is intended as a text for senior/graduate level college students with emphasis
on real-time DSP implementations and applications. This book can also serve as a
desktop reference for practicing engineer and embedded system programmer to learn
DSP concepts and to develop real-time DSP applications at work. We use a practical
approach that avoids a lot of theoretical derivations. Many useful DSP textbooks with
solid mathematical proofs are listed at the end of each chapter. To efficiently develop a
DSP system, the reader must understand DSP algorithms as well as basic DSP chip
architecture and programming. It is helpful to have several manuals and application
notes on the TMS320C55x from Texas Instruments at http://www.ti.com.
The DSP processor we will use as an example in this book is the TMS320C55x, the
newest 16-bit fixed-point DSP processor fromTexas Instruments. To effectively illustrate
real-time DSP concepts and applications, MATLAB will be introduced for analysis and
filter design, C will be used for implementing DSP algorithms, and Code Composer
Studio (CCS) of the TMS320C55x are integrated into lab experiments, projects, and
applications. To efficiently utilize the advanced DSP architecture for fast software
development and maintenance, the mixing of C and assembly programs are emphasized.
Chapter 1 reviews the fundamentals of real-time DSP functional blocks, DSP hard-
ware options, fixed- and floating-point DSP devices, real-time constraints, algorithm
development, selection of DSP chips, and software development. In Chapter 2, we
introduce the architecture and assembly programming of the TMS320C55x. Chapter
3 presents some fundamental DSP concepts in time domain and practical considerations
for the implementation of digital filters and algorithms on DSP hardware. Readers who
are familiar with these DSP fundamentals should be able to skip through some of these
sections. However, most notations used throughout the book will be defined in this
chapter. In Chapter 4, the Fourier series, the Fourier transform, the z-transform, and
the discrete Fourier transforms are introduced. Frequency analysis is extremely helpful
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
in understanding the characteristics of both signals and systems. Chapter 5 is focused on
the design, implementation, and application of FIR filters; digital IIR filters are covered
in Chapter 6, and adaptive filters are presented in Chapter 8. The development,
implementation, and application of FFT algorithms are introduced in Chapter 7. In
Chapter 9, we introduce some selected DSP applications in communications that have
played an important role in the realization of the systems.
As with any book attempting to capture the state of the art at a given time, there will
necessarily be omissions that are necessitated by the rapidly evolving developments in
this dynamic field of exciting practical interest. We hope, at least, that this book will
serve as a guide for what has already come and as an inspiration for what will follow. To
aid teaching of the course a Solution Manual that presents detailed solutions to most of
the problems in the book is available from the publisher.
Availability of Software
The MATLAB, C, and assembly programs that implement many DSP examples and
applications are listed in the book. These programs along with many other programs
for DSP implementations and lab experiments are available in the software package
at http://www.ceet.niu.edu/faculty/kuo/books/rtdsp.html and http://pages.prodigy.net/
sunheel/web/dspweb.htm. Several real-world data files for some applications introduced
in the book also are included in the software package. The list of files in the software
package is given in Appendix D. It is not critical you have this software as you read the
book, but it will helpyoutogaininsight intothe implementationof DSPalgorithms, andit
will be required for doing experiments at the last section of each chapter. Some of these
experiments involve minor modification of the example code. By examining, studying and
modifyingtheexamplecode, thesoftwarecanalsobeusedas aprototypefor other practical
applications. Everyattempt has beenmade toensure the correctness of the code. We would
appreciate readers bringing to our attention (kuo@ceet.niu.edu) any coding errors so that
we can correct and update the codes available in the software package on the web.
Acknowledgments
We are grateful to Maria Ho and Christina Peterson at Texas Instruments, and Naomi
Fernandes at Math Works, who provided the necessary support to write the book in a
short period. The first author thanks manyof his students whohave takenhis DSPcourses,
Senior Design Projects, and Master Thesis courses. He is indebted to Gene Frentz, Dr.
Qun S. Lin, and Dr. Panos Papamichalis of Texas Instruments, John Kronenburger of
Tellabs, and Santo LaMantia of Shure Brothers, for their support of DSP activities at
Northern Illinois University. He also thanks Jennifer Y. Kuo for the proofreading of the
book. The secondauthor wishes tothankRobert DeNardo, DavidBaughman, andChuck
Brokish of Texas Instruments, for their valuable inputs, help, and encouragement during
the course of writing this book. We would like to thank Peter Mitchell, editor at Wiley, for
his support of this project. We also like to thank the staff at Wiley for the final preparation
of the book. Finally, we thank our parents and families for their endless love, encourage-
ment, and the understanding they have shown during the whole time.
Sen M. Kuo and Bob H. Lee
xvi PREFACE
1
Introduction to Real-Time
Digital Signal Processing
Signals can be divided into three categories ± continuous-time (analog) signals,
discrete-time signals, and digital signals. The signals that we encounter daily are mostly
analog signals. These signals are defined continuously in time, have an infinite range
of amplitude values, and can be processed using electrical devices containing both
active and passive circuit elements. Discrete-time signals are defined only at a particular
set of time instances. Therefore they can be represented as a sequence of numbers that
have a continuous range of values. On the other hand, digital signals have discrete
values in both time and amplitude. In this book, we design and implement digital
systems for processing digital signals using digital hardware. However, the analysis
of such signals and systems usually uses discrete-time signals and systems for math-
ematical convenience. Therefore we use the term `discrete-time' and `digital' inter-
changeably.
Digital signal processing (DSP) is concerned with the digital representation of signals
and the use of digital hardware to analyze, modify, or extract information from these
signals. The rapid advancement in digital technology in recent years has created the
implementation of sophisticated DSP algorithms that make real-time tasks feasible. A
great deal of research has been conducted to develop DSP algorithms and applications.
DSP is now used not only in areas where analog methods were used previously, but also
in areas where applying analog techniques is difficult or impossible.
There are many advantages in using digital techniques for signal processing rather
than traditional analog devices (such as amplifiers, modulators, and filters). Some of the
advantages of a DSP system over analog circuitry are summarized as follows:
1. Flexibility. Functions of a DSP system can be easily modified and upgraded with
software that has implemented the specific algorithm for using the same hardware.
One can design a DSP system that can be programmed to perform a wide variety of
tasks by executing different software modules. For example, a digital camera may
be easily updated (reprogrammed) from using JPEG ( joint photographic experts
group) image processing to a higher quality JPEG2000 image without actually
changing the hardware. In an analog system, however, the whole circuit design
would need to be changed.
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
2. Reproducibility. The performance of a DSP system can be repeated precisely from
one unit to another. This is because the signal processing of DSP systems work
directly with binary sequences. Analog circuits will not perform as well from each
circuit, even if they are built following identical specifications, due to component
tolerances in analog components. In addition, by using DSP techniques, a digital
signal can be transferred or reproduced many times without degrading its signal
quality.
3. Reliability. The memory and logic of DSP hardware does not deteriorate with
age. Therefore the field performance of DSP systems will not drift with changing
environmental conditions or aged electronic components as their analog counter-
parts do. However, the data size (wordlength) determines the accuracy of a DSP
system. Thus the system performance might be different from the theoretical expect-
ation.
4. Complexity. Using DSP allows sophisticated applications such as speech or image
recognition to be implemented for lightweight and low power portable devices. This
is impractical using traditional analog techniques. Furthermore, there are some
important signal processing algorithms that rely on DSP, such as error correcting
codes, data transmission and storage, data compression, perfect linear phase filters,
etc., which can barely be performed by analog systems.
With the rapid evolution in semiconductor technology in the past several years, DSP
systems have a lower overall cost compared to analog systems. DSP algorithms can be
developed, analyzed, and simulated using high-level language and software tools such as
C=C÷÷ and MATLAB (matrix laboratory). The performance of the algorithms can be
verified using a low-cost general-purpose computer such as a personal computer (PC).
Therefore a DSP system is relatively easy to develop, analyze, simulate, and test.
There are limitations, however. For example, the bandwidth of a DSP system is
limited by the sampling rate and hardware peripherals. The initial design cost of a
DSP system may be expensive, especially when large bandwidth signals are involved.
For real-time applications, DSP algorithms are implemented using a fixed number of
bits, which results in a limited dynamic range and produces quantization and arithmetic
errors.
1.1 Basic Elements of Real-Time DSP Systems
There are two types of DSP applications ± non-real-time and real time. Non-real-time
signal processing involves manipulating signals that have already been collected and
digitized. This may or may not represent a current action and the need for the result
is not a function of real time. Real-time signal processing places stringent demands
on DSP hardware and software design to complete predefined tasks within a certain
time frame. This chapter reviews the fundamental functional blocks of real-time DSP
systems.
The basic functional blocks of DSP systems are illustrated in Figure 1.1, where a real-
world analog signal is converted to a digital signal, processed by DSP hardware in
2 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
Other digital
systems
Anti-aliasing
filter
ADC
x(n)
DSP
hardware
Other digital
systems
DAC
Reconstruction
filter
y(n)
x(t) xЈ(t)
Amplifier
Amplifier
y(t) yЈ(t)
Input channels
Output channels
Figure 1.1 Basic functional blocks of real-time DSP system
digital form, and converted back into an analog signal. Each of the functional blocks in
Figure 1.1 will be introduced in the subsequent sections. For some real-time applica-
tions, the input data may already be in digital form and/or the output data may not need
to be converted to an analog signal. For example, the processed digital information may
be stored in computer memory for later use, or it may be displayed graphically. In other
applications, the DSP system may be required to generate signals digitally, such as
speech synthesis used for cellular phones or pseudo-random number generators for
CDMA (code division multiple access) systems.
1.2 Input and Output Channels
In this book, a time-domain signal is denoted with a lowercase letter. For example, x(t)
in Figure 1.1 is used to name an analog signal of x with a relationship to time t. The time
variable t takes on a continuum of values between ÷· and ·. For this reason we say
x(t) is a continuous-time signal. In this section, we first discuss how to convert analog
signals into digital signals so that they can be processed using DSP hardware. The
process of changing an analog signal to a xdigital signal is called analog-to-digital (A/D)
conversion. An A/D converter (ADC) is usually used to perform the signal conversion.
Once the input digital signal has been processed by the DSP device, the result, y(n), is
still in digital form, as shown in Figure 1.1. In many DSP applications, we need to
reconstruct the analog signal after the digital processing stage. In other words, we must
convert the digital signal y(n) back to the analog signal y(t) before it is passed to an
appropriate device. This process is called the digital-to-analog (D/A) conversion, typi-
cally performed by a D/A converter (DAC). One example would be CD (compact disk)
players, for which the music is in a digital form. The CD players reconstruct the analog
waveform that we listen to. Because of the complexity of sampling and synchronization
processes, the cost of an ADC is usually considerably higher than that of a DAC.
1.2.1 Input Signal Conditioning
As shown in Figure 1.1, the analog signal, x
/
(t), is picked up by an appropriate
electronic sensor that converts pressure, temperature, or sound into electrical signals.
INPUT AND OUTPUT CHANNELS 3
For example, a microphone can be used to pick up sound signals. The sensor output,
x
/
(t), is amplified by an amplifier with gain value g. The amplified signal is
x(t) = gx
/
(t): (1:2:1)
The gain value g is determined such that x(t) has a dynamic range that matches the
ADC. For example, if the peak-to-peak range of the ADC is ±5 volts (V), then g may be
set so that the amplitude of signal x(t) to the ADC is scaled between ±5V. In practice, it
is very difficult to set an appropriate fixed gain because the level of x
/
(t) may be
unknown and changing with time, especially for signals with a larger dynamic range
such as speech. Therefore an automatic gain controller (AGC) with time-varying gain
determined by DSP hardware can be used to effectively solve this problem.
1.2.2 A/D Conversion
As shown in Figure 1.1, the ADC converts the analog signal x(t) into the digital signal
sequence x(n). Analog-to-digital conversion, commonly referred as digitization, consists
of the sampling and quantization processes as illustrated in Figure 1.2. The sampling
process depicts a continuously varying analog signal as a sequence of values. The basic
sampling function can be done with a `sample and hold' circuit, which maintains the
sampled level until the next sample is taken. Quantization process approximates a
waveform by assigning an actual number for each sample. Therefore an ADC consists
of two functional blocks ± an ideal sampler (sample and hold) and a quantizer (includ-
ing an encoder). Analog-to-digital conversion carries out the following steps:
1. The bandlimited signal x(t) is sampled at uniformly spaced instants of time, nT,
where n is a positive integer, and T is the sampling period in seconds. This sampling
process converts an analog signal into a discrete-time signal, x(nT), with continuous
amplitude value.
2. The amplitude of each discrete-time sample is quantized into one of the 2
B
levels,
where B is the number of bits the ADC has to represent for each sample. The
discrete amplitude levels are represented (or encoded) into distinct binary words
x(n) with a fixed wordlength B. This binary sequence, x(n), is the digital signal for
DSP hardware.
x(t)
Ideal sampler
x(nT)
Quantizer
x(n)
A/D converter
Figure 1.2 Block diagram of A/D converter
4 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
The reason for making this distinction is that each process introduces different distor-
tions. The sampling process brings in aliasing or folding distortions, while the encoding
process results in quantization noise.
1.2.3 Sampling
An ideal sampler can be considered as a switch that is periodically open and closed every
T seconds and
T =
1
f
s
, (1:2:2)
where f
s
is the sampling frequency (or sampling rate) in hertz (Hz, or cycles per
second). The intermediate signal, x(nT), is a discrete-time signal with a continuous-
value (a number has infinite precision) at discrete time nT, n = 0, 1, . . ., ·as illustrated
in Figure 1.3. The signal x(nT) is an impulse train with values equal to the amplitude
of x(t) at time nT. The analog input signal x(t) is continuous in both time and
amplitude. The sampled signal x(nT) is continuous in amplitude, but it is defined
only at discrete points in time. Thus the signal is zero except at the sampling instants
t = nT.
In order to represent an analog signal x(t) by a discrete-time signal x(nT) accurately,
two conditions must be met:
1. The analog signal, x(t), must be bandlimited by the bandwidth of the signal f
M
.
2. The sampling frequency, f
s
, must be at least twice the maximum frequency com-
ponent f
M
in the analog signal x(t). That is,
f
s
_ 2 f
M
: (1:2:3)
This is Shannon's sampling theorem. It states that when the sampling frequency is
greater than twice the highest frequency component contained in the analog signal, the
original signal x(t) can be perfectly reconstructed from the discrete-time sample x(nT).
The sampling theorem provides a basis for relating a continuous-time signal x(t) with
Time, t
x(nT)
0 T 2T 3T 4T
x(t)
Figure 1.3 Example of analog signal x(t) and discrete-time signal x(nT)
INPUT AND OUTPUT CHANNELS 5
the discrete-time signal x(nT) obtained from the values of x(t) taken T seconds apart. It
also provides the underlying theory for relating operations performed on the sequence
to equivalent operations on the signal x(t) directly.
The minimum sampling frequency f
s
= 2f
M
is the Nyquist rate, while f
N
= f
s
=2 is
the Nyquist frequency (or folding frequency). The frequency interval [÷f
s
=2, f
s
=2[
is called the Nyquist interval. When an analog signal is sampled at sampling frequency,
f
s
, frequency components higher than f
s
=2 fold back into the frequency range [0, f
s
=2[.
This undesired effect is known as aliasing. That is, when a signal is sampled
perversely to the sampling theorem, image frequencies are folded back into the desired
frequency band. Therefore the original analog signal cannot be recovered from the
sampled data. This undesired distortion could be clearly explained in the frequency
domain, which will be discussed in Chapter 4. Another potential degradation is due to
timing jitters on the sampling pulses for the ADC. This can be negligible if a higher
precision clock is used.
For most practical applications, the incoming analog signal x(t) may not be band-
limited. Thus the signal has significant energies outside the highest frequency of
interest, and may contain noise with a wide bandwidth. In other cases, the sampling
rate may be pre-determined for a given application. For example, most voice commu-
nication systems use an 8 kHz (kilohertz) sampling rate. Unfortunately, the maximum
frequency component in a speech signal is much higher than 4 kHz. Out-of-band signal
components at the input of an ADC can become in-band signals after conversion
because of the folding over of the spectrum of signals and distortions in the discrete
domain. To guarantee that the sampling theorem defined in Equation (1.2.3) can be
fulfilled, an anti-aliasing filter is used to band-limit the input signal. The anti-aliasing
filter is an analog lowpass filter with the cut-off frequency of
f
c
_
f
s
2
: (1:2:4)
Ideally, an anti-aliasing filter should remove all frequency components above the
Nyquist frequency. In many practical systems, a bandpass filter is preferred in order
to prevent undesired DC offset, 60 Hz hum, or other low frequency noises. For example,
a bandpass filter with passband from 300 Hz to 3200 Hz is used in most telecommunica-
tion systems.
Since anti-aliasing filters used in real applications are not ideal filters, they cannot
completely remove all frequency components outside the Nyquist interval. Any fre-
quency components and noises beyond half of the sampling rate will alias into the
desired band. In addition, since the phase response of the filter may not be linear, the
components of the desired signal will be shifted in phase by amounts not proportional to
their frequencies. In general, the steeper the roll-off, the worse the phase distortion
introduced by a filter. To accommodate practical specifications for anti-aliasing filters,
the sampling rate must be higher than the minimum Nyquist rate. This technique is
known as oversampling. When a higher sampling rate is used, a simple low-cost anti-
aliasing filter with minimum phase distortion can be used.
Example 1.1: Given a sampling rate for a specific application, the sampling period
can be determined by (1.2.2).
6 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
(a) In narrowband telecommunication systems, the sampling rate f
s
= 8 kHz,
thus the sampling period T = 1=8 000 seconds = 125 ms (microseconds).
Note that 1 ms = 10
÷6
seconds.
(b) In wideband telecommunication systems, the sampling is given as
f
s
= 16 kHz, thus T = 1=16 000 seconds = 62:5 ms.
(c) In audio CDs, the sampling rate is f
s
= 44:1 kHz, thus T = 1=44 100 seconds
= 22:676 ms.
(d) In professional audio systems, the sampling rate f
s
= 48 kHz, thus
T = 1=48 000 seconds = 20:833 ms.
1.2.4 Quantizing and Encoding
In the previous sections, we assumed that the sample values x(nT) are represented
exactly with infinite precision. An obvious constraint of physically realizable digital
systems is that sample values can only be represented by a finite number of bits.
The fundamental distinction between discrete-time signal processing and DSP is the
wordlength. The former assumes that discrete-time signal values x(nT) have infinite
wordlength, while the latter assumes that digital signal values x(n) only have a limited
B-bit.
We now discuss a method of representing the sampled discrete-time signal x(nT) as a
binary number that can be processed with DSP hardware. This is the quantizing and
encoding process. As shown in Figure 1.3, the discrete-time signal x(nT) has an analog
amplitude (infinite precision) at time t = nT. To process or store this signal with DSP
hardware, the discrete-time signal must be quantized to a digital signal x(n) with a finite
number of bits. If the wordlength of an ADC is B bits, there are 2
B
different values
(levels) that can be used to represent a sample. The entire continuous amplitude range is
divided into 2
B
subranges. Amplitudes of waveform that are in the same subrange are
assigned the same amplitude values. Therefore quantization is a process that represents
an analog-valued sample x(nT) with its nearest level that corresponds to the digital
signal x(n). The discrete-time signal x(nT) is a sequence of real numbers using infinite
bits, while the digital signal x(n) represents each sample value by a finite number of bits
which can be stored and processed using DSP hardware.
The quantization process introduces errors that cannot be removed. For example, we
can use two bits to define four equally spaced levels (00, 01, 10, and 11) to classify the
signal into the four subranges as illustrated in Figure 1.4. In this figure, the symbol `o'
represents the discrete-time signal x(nT), and the symbol `v' represents the digital signal
x(n).
In Figure 1.4, the difference between the quantized number and the original value is
defined as the quantization error, which appears as noise in the output. It is also called
quantization noise. The quantization noise is assumed to be random variables that are
uniformly distributed in the intervals of quantization levels. If a B-bit quantizer is used,
the signal-to-quantization-noise ratio (SNR) is approximated by (will be derived in
Chapter 3)
INPUT AND OUTPUT CHANNELS 7
T 2T 3T
00
01
10
11
Quantization level
Time, t
x(t)
0
Quantization errors
Figure 1.4 Digital samples using a 2-bit quantizer
SNR ~ 6BdB: (1:2:5)
This is a theoretical maximum. When real input signals and converters are used, the
achievable SNR will be less than this value due to imperfections in the fabrication of
A/D converters. As a result, the effective number of bits may be less than the number
of bits in the ADC. However, Equation (1.2.5) provides a simple guideline for determin-
ing the required bits for a given application. For each additional bit, a digital signal has
about a 6-dB gain in SNR. For example, a 16-bit ADC provides about 96 dB SNR. The
more bits used to represent a waveform sample, the smaller the quantization noise will
be. If we had an input signal that varied between 0 and 5 V, using a 12-bit ADC, which
has 4096 (2
12
) levels, the least significant bit (LSB) would correspond to 1.22 mV
resolution. An 8-bit ADC with 256 levels can only provide up to 19.5 mV resolution.
Obviously with more quantization levels, one can represent the analog signal more
accurately. The problems of quantization and their solutions will be further discussed in
Chapter 3.
If the uniform quantization scheme shown in Figure 1.4 can adequately represent
loud sounds, most of the softer sounds may be pushed into the same small value. This
means soft sounds may not be distinguishable. To solve this problem, a quantizer whose
quantization step size varies according to the signal amplitude can be used. In practice,
the non-uniform quantizer uses a uniform step size, but the input signal is compressed
first. The overall effect is identical to the non-uniform quantization. For example, the
logarithm-scaled input signal, rather than the input signal itself, will be quantized. After
processing, the signal is reconstructed at the output by expanding it. The process of
compression and expansion is called companding (compressing and expanding). For
example, the m-law (used in North America and parts of Northeast Asia) and A-law
(used in Europe and most of the rest of the world) companding schemes are used in most
digital communications.
As shown in Figure 1.1, the input signal to DSP hardware may be a digital signal
from other DSP systems. In this case, the sampling rate of digital signals from other
digital systems must be known. The signal processing techniques called interpolation or
decimation can be used to increase or decrease the existing digital signals' sampling
rates. Sampling rate changes are useful in many applications such as interconnecting
DSP systems operating at different rates. A multirate DSP system uses more than one
sampling frequency to perform its tasks.
8 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
1.2.5 D/A Conversion
Most commercial DACs are zero-order-hold, which means they convert the binary
input to the analog level and then simply hold that value for T seconds until the next
sampling instant. Therefore the DAC produces a staircase shape analog waveform y
/
(t),
which is shown as a solid line in Figure 1.5. The reconstruction (anti-imaging and
smoothing) filter shown in Figure 1.1 smoothes the staircase-like output signal gener-
ated by the DAC. This analog lowpass filter may be the same as the anti-aliasing filter
with cut-off frequency f
c
_ f
s
=2, which has the effect of rounding off the corners of the
staircase signal and making it smoother, which is shown as a dotted line in Figure 1.5.
High quality DSP applications, such as professional digital audio, require the use of
reconstruction filters with very stringent specifications.
From the frequency-domain viewpoint (will be presented in Chapter 4), the output of
the DAC contains unwanted high frequency or image components centered at multiples
of the sampling frequency. Depending on the application, these high-frequency compon-
ents may cause undesired side effects. Take an audio CD player for example. Although
the image frequencies may not be audible, they could overload the amplifier and cause
inter-modulation with the desired baseband frequency components. The result is an
unacceptable degradation in audio signal quality.
The ideal reconstruction filter has a flat magnitude response and linear phase in the
passband extending from the DC to its cut-off frequency and infinite attenuation in
the stopband. The roll-off requirements of the reconstruction filter are similar to those
of the anti-aliasing filter. In practice, switched capacitor filters are preferred because of
their programmable cut-off frequency and physical compactness.
1.2.6 Input/Output Devices
There are two basic ways of connecting A/D and D/A converters to DSP devices: serial
and parallel. A parallel converter receives or transmits all the B bits in one pass, while
the serial converters receive or transmit B bits in a serial data stream. Converters with
parallel input and output ports must be attached to the DSP's address and data buses,
yЈ(t)
Time, t
0 T 2T 3T 4T 5T
Smoothed output
signal
Figure 1.5 Staircase waveform generated by a DAC
INPUT AND OUTPUT CHANNELS 9
which are also attached to many different types of devices. With different memory
devices (RAM, EPROM, EEPROM, or flash memory) at different speeds hanging on
DSP's data bus, driving the bus may become a problem. Serial converters can be
connected directly to the built-in serial ports of DSP devices. This is why many practical
DSP systems use serial ADCs and DACs.
Many applications use a single-chip device called an analog interface chip (AIC) or
coder/decoder (CODEC), which integrates an anti-aliasing filter, an ADC, a DAC, and a
reconstruction filter all on a single piece of silicon. Typical applications include modems,
speech systems, and industrial controllers. Many standards that specify the nature of the
CODEC have evolved for the purposes of switching and transmission. These devices
usually use a logarithmic quantizer, i.e., A-law or m-law, which must be converted into a
linear format for processing. The availability of inexpensive companded CODEC justi-
fies their use as front-end devices for DSP systems. DSP chips implement this format
conversion in hardware or in software by using a table lookup or calculation.
The most popular commercially available ADCs are successive approximation, dual
slope, flash, and sigma-delta. The successive-approximation ADC produces a B-bit
output in B cycles of its clock by comparing the input waveform with the output of a
digital-to-analog converter. This device uses a successive-approximation register to split
the voltage range in half in order to determine where the input signal lies. According to
the comparator result, one bit will be set or reset each time. This process proceeds
from the most significant bit (MSB) to the LSB. The successive-approximation type of
ADC is generally accurate and fast at a relatively low cost. However, its ability to follow
changes in the input signal is limited by its internal clock rate, so that it may be slow to
respond to sudden changes in the input signal.
The dual-slope ADC uses an integrator connected to the input voltage and a reference
voltage. The integrator starts at zero condition, and it is charged for a limited time. The
integrator is then switched to a known negative reference voltage and charged in the
opposite direction until it reaches zero volts again. At the same time, a digital counter
starts to record the clock cycles. The number of counts required for the integrator
output voltage to get back to zero is directly proportional to the input voltage. This
technique is very precise and can produce ADCs with high resolution. Since the
integrator is used for input and reference voltages, any small variations in temperature
and aging of components have little or no effect on these types of converters. However,
they are very slow and generally cost more than successive-approximation ADCs.
A voltage divider made by resistors is used to set reference voltages at the flash ADC
inputs. The major advantage of a flash ADC is its speed of conversion, which is simply
the propagation delay time of the comparators. Unfortunately, a B-bit ADC needs
(2
B
÷1) comparators and laser-trimmed resistors. Therefore commercially available
flash ADCs usually have lower bits.
The block diagram of a sigma±delta ADC is illustrated in Figure 1.6. Sigma±delta
ADCs use a 1-bit quantizer with a very high sampling rate. Thus the requirements for an
anti-aliasing filter are significantly relaxed (i.e., the lower roll-off rate and smaller flat
response in passband). In the process of quantization, the resulting noise power is spread
evenly over the entire spectrum. As a result, the noise power within the band of interest is
lower. In order to match the output frequency with the systemand increase its resolution,
a decimator is used. The advantages of the sigma±delta ADCs are high resolution and
good noise characteristics at a competitive price because they use digital filters.
10 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
Analog
input

Σ
Sigma
Delta
1-bit B-bit
1-bit
DAC
1-bit
ADC

+
Digital
decimator
Figure 1.6 A block of a sigma-delta ADC
1.3 DSP Hardware
DSP systems require intensive arithmetic operations, especially multiplication and
addition. In this section, different digital hardware architectures for DSP applications
will be discussed.
1.3.1 DSP Hardware Options
As shown in Figure 1.1, the processing of the digital signal x(n) is carried out using the
DSP hardware. Although it is possible to implement DSP algorithms on any digital
computer, the throughput (processing rate) determines the optimum hardware plat-
form. Four DSP hardware platforms are widely used for DSP applications:
1. general-purpose microprocessors and microcontrollers (mP),
2. general-purpose digital signal processors (DSP chips),
3. digital building blocks (DBB) such as multiplier, adder, program controller, and
4. special-purpose (custom) devices such as application specific integrated circuits
(ASIC).
The hardware characteristics are summarized in Table 1.1.
ASIC devices are usually designed for specific tasks that require a lot of DSP MIPS
(million instructions per second), such as fast Fourier transform (FFT) devices and
Reed±Solomon coders used by digital subscriber loop (xDSL) modems. These devices
are able to perform their limited functions much faster than general-purpose DSP chips
because of their dedicated architecture. These application-specific products enable the
use of high-speed functions optimized in hardware, but they lack the programmability
to modify the algorithm, so they are suitable for implementing well-defined and well-
tested DSP algorithms. Therefore applications demanding high speeds typically employ
ASICs, which allow critical DSP functions to be implemented in the hardware. The
availability of core modules for common DSP functions has simplified the ASIC design
tasks, but the cost of prototyping an ASIC device, a longer design cycle, insufficient
DSP HARDWARE 11
Table 1.1 Summary of DSP hardware implementations
ASIC DBB mP DSP chips
Chip count 1 > 1 1 1
Flexibility none limited programmable programmable
Design time long medium short short
Power consumption low medium±high medium low±medium
Processing speed high high low±medium medium±high
Reliability high low±medium high high
Development cost high medium low low
Production cost low high low±medium low±medium
Processor
Address bus
Data bus
Memory
Processor
Address bus 1
Address bus 2
Data bus 1
Data bus 2
Memory 1 Memory 2
(a) (b)
Figure 1.7 Different memory architectures: (a) Harvard architecture, and (b) von Newmann
architecture
standard development tools support, and the lack of reprogramming flexibility some-
times outweigh their benefits.
Digital building blocks offer a more general-purpose approach to high-speed DSP
design. These components, including multipliers, arithmetic logic units (ALUs), sequen-
cers, etc., are joined together to build a custom DSP architecture for a specific applica-
tion. Performance can be significantly higher than general-purpose DSP devices.
However, the disadvantages are similar to those of the special-purpose DSP devices ±
lack of standard design tools, extended design cycles, and high component cost.
General architectures for computers and microprocessors fall into two categories:
Harvard architecture and von Neumann architecture. Harvard architecture has a
separate memory space for the program and the data, so that both memories can be
accessed simultaneously, see Figure 1.7(a). The von Neumann architecture assumes that
12 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
there is no intrinsic difference between the instructions and the data, and that the
instructions can be partitioned into two major fields containing the operation command
and the address of the operand. Figure 1.7(b) shows the memory architecture of the von
Neumann model. Most general-purpose microprocessors use the von Neumann archi-
tecture. Operations such as add, move, and subtract are easy to perform. However,
complex instructions such as multiplication and division are slow since they need a
series of shift, addition, or subtraction operations. These devices do not have the
architecture or the on-chip facilities required for efficient DSP operations. They may
be used when a small amount of signal processing work is required in a much larger
system. Their real-time DSP performance does not compare well with even the cheaper
general-purpose DSP devices, and they would not be a cost-effective solution for many
DSP tasks.
A DSP chip (digital signal processor) is basically a microprocessor whose architecture
is optimized for processing specific operations at high rates. DSP chips with architec-
tures and instruction sets specifically designed for DSP applications have been launched
by Texas Instruments, Motorola, Lucent Technologies, Analog Devices, and many
other companies. The rapid growth and the exploitation of DSP semiconductor tech-
nology are not a surprise, considering the commercial advantages in terms of the fast,
flexible, and potentially low-cost design capabilities offered by these devices. General-
purpose-programmable DSP chip developments are supported by software develop-
ment tools such as C compilers, assemblers, optimizers, linkers, debuggers, simulators,
and emulators. Texas Instruments' TMS320C55x, a programmable, high efficiency, and
ultra low-power DSP chip, will be discussed in the next chapter.
1.3.2 Fixed- and Floating-Point Devices
A basic distinction between DSP chips is their fixed-point or floating-point architectures.
The fixed-point representation of signals and arithmetic will be discussed in Chapter 3.
Fixed-point processors are either 16-bit or 24-bit devices, while floating-point processors
are usually 32-bit devices. A typical 16-bit fixed-point processor, such as the
TMS320C55x, stores numbers in a 16-bit integer format. Although coefficients and
signals are only stored with 16-bit precision, intermediate values (products) may be kept
at 32-bit precision within the internal accumulators in order to reduce cumulative round-
ing errors. Fixed-point DSP devices are usually cheaper and faster than their floating-
point counterparts because they use less silicon and have fewer external pins.
A typical 32-bit floating-point DSP device, such as the TMS320C3x, stores a 24-bit
mantissa and an 8-bit exponent. A 32-bit floating-point format gives a large dynamic
range. However, the resolution is still only 24 bits. Dynamic range limitations may be
virtually ignored in a design using floating-point DSP chips. This is in contrast to fixed-
point designs, where the designer has to apply scaling factors to prevent arithmetic
overflow, which is a very difficult and time-consuming process.
Floating-point devices may be needed in applications where coefficients vary in time,
signals and coefficients have a large dynamic range, or where large memory structures
are required, such as in image processing. Other cases where floating-point devices can
be justified are where development costs are high and production volumes are low. The
faster development cycle for a floating-point device may easily outweigh the extra cost
DSP HARDWARE 13
of the DSP device itself. Floating-point DSP chips also allow the efficient use of
the high-level C compilers and reduce the need to identify the system's dynamic range.
1.3.3 Real-Time Constraints
A limitation of DSP systems for real-time applications is that the bandwidth of the
system is limited by the sampling rate. The processing speed determines the rate at
which the analog signal can be sampled. For example, a real-time DSP system demands
that the signal processing time, t
p
, must be less than the sampling period, T, in order to
complete the processing task before the new sample comes in. That is,
t
p
< T: (1:3:1)
This real-time constraint limits the highest frequency signal that can be processed by a
DSP system. This is given as
f
M
_
f
s
2
<
1
2t
p
: (1:3:2)
It is clear that the longer the processing time t
p
, the lower the signal bandwidth f
M
.
Although new and faster DSP devices are introduced, there is still a limit to the
processing that can be done in real time. This limit becomes even more apparent when
system cost is taken into consideration. Generally, the real-time bandwidth can be
increased by using faster DSP chips, simplified DSP algorithms, optimized DSP pro-
grams, and parallel processing using multiple DSP chips, etc. However, there is still a
trade-off between costs and system performances, with many applications simply not
being economical at present.
1.4 DSP System Design
A generalized DSP system design is illustrated in Figure 1.8. For a given application, the
theoretical aspects of DSP system specifications such as system requirements, signal
analysis, resource analysis, and configuration analysis are first performed to define the
system requirements.
1.4.1 Algorithm Development
The algorithm for a given application is initially described using difference equations or
signal-flow block diagrams with symbolic names for the inputs and outputs. In docu-
menting the algorithm, it is sometimes helpful to further clarify which inputs and
outputs are involved by means of a data flow diagram. The next stage of the develop-
ment process is to provide more details on the sequence of operations that must be
performed in order to derive the output from the input. There are two methods for
characterizing the sequence of steps in a program: flowcharts or structured descriptions.
14 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
H
A
R
D
W
A
R
E
S
O
F
T
W
A
R
E
System requirements specifications
Algorithm development and simulation
Select DSP devices
System integration and debug
System testing and release
Application
Software
architecture
Coding and
Debugging
Hardware
Schematic
Hardware
Prototype
Figure 1.8 Simplified DSP system design flow
MATLAB or C/C++
ADC
Other
computers
Other
computers
DAC
Signal generators
DSP
algorithms
DSP
software
Data
files
Data
files
Figure 1.9 DSP software development using a general-purpose computer
At the algorithm development stage, we most likely work with high-level DSP tools
(such as MATLAB or C=C÷÷) that enable algorithmic-level system simulations. We
then migrate the algorithm to software, hardware, or both, depending on our specific
needs. A DSP application or algorithm can be first simulated using a general-purpose
computer, such as a PC, so that it can be analyzed and tested off-line using simulated
input data. A block diagram of general-purpose computer implementation is illustrated
in Figure 1.9. The test signals may be internally generated by signal generators or
digitized from an experimental setup based on the given application. The program uses
the stored signal samples in data file(s) as input(s) to produce output signals that will be
saved in data file(s).
DSP SYSTEM DESIGN 15
Advantages of developing DSP software on a general-purpose computer are:
1. Using the high-level languages such as MATLAB, C=C÷÷, or other DSP software
packages can significantly save algorithm and software development time. In add-
ition, C programs are portable to different DSP hardware platforms.
2. It is easy to debug and modify programs.
3. Input/output operations based on disk files are simple to implement and the behav-
iors of the system are easy to analyze.
4. Using the floating-point data format can achieve higher precision.
5. With fixed-point simulation, bit-true verification of an algorithm against fixed-
point DSP implementation can easily be compared.
1.4.2 Selection of DSP Chips
A choice of DSP chip from many available devices requires a full understanding of the
processing requirements of the DSP system under design. The objective is to select the
device that meets the project's time-scales and provides the most cost-effective solution.
Some decisions can be made at an early stage based on computational power, resolu-
tion, cost, etc. In real-time DSP, the efficient flow of data into and out of the processor
is also critical. However, these criteria will probably still leave a number of candidate
devices for further analysis. For high-volume applications, the cheapest device that can
do the job should be chosen. For low- to medium-volume applications, there will be a
trade-off between development time, development tool cost, and the cost of the DSP
device itself. The likelihood of having higher-performance devices with upwards-
compatible software in the future is also an important factor.
When processing speed is at a premium, the only valid comparison between devices is
on an algorithm-implementation basis. Optimum code must be written for both devices
and then the execution time must be compared. Other important factors are memory size
and peripheral devices, such as serial and parallel interfaces, which are available on-chip.
In addition, a full set of development tools and supports are important for DSP chip
selection, including:
1. Software development tools, such as assemblers, linkers, simulators, and C com-
pilers.
2. Commercially available DSP boards for software development and testing before
the target DSP hardware is available.
3. Hardware testing tools, such as in-circuit emulators and logic analyzers.
4. Development assistance, such as application notes, application libraries, data
books, real-time debugging hardware, low-cost prototyping, etc.
16 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
1.4.3 Software Development
The four common measures of good DSP software are reliability, maintainability,
extensibility, and efficiency. A reliable program is one that seldom (or never) fails.
Since most programs will occasionally fail, a maintainable program is one that is easy
to fix. A truly maintainable program is one that can be fixed by someone other than
the original programmer. In order for a program to be truly maintainable, it must be
portable on more than one type of hardware. An extensible program is one that can
be easily modified when the requirements change, new functions need to be added, or
new hardware features need to be exploited. An efficient DSP program will use the
processing capabilities of the target hardware to minimize execution time.
A program is usually tested in a finite number of ways much smaller than the number
of input data conditions. This means that a program can be considered reliable only
after years of bug-free use in many different environments. A good DSP program often
contains many small functions with only one purpose, which can be easily reused by
other programs for different purposes. Programming tricks should be avoided at all
costs as they will often not be reliable and will almost always be difficult for someone
else to understand even with lots of comments. In addition, use variable names that are
meaningful in the context of the program.
As shown in Figure 1.8, the hardware and software design can be conducted at the
same time for a given DSP application. Since there is a lot of interdependence factors
between hardware and software, the ideal DSP designer will be a true `system' engineer,
capable of understanding issues with both hardware and software. The cost of hardware
has gone down dramatically in recent years. The majority of the cost of a DSP solution
now resides in software development. This section discussed some issues regarding
software development.
The software life cycle involves the completion of a software project: the project
definition, the detailed specification, coding and modular testing, integration, and
maintenance. Software maintenance is a significant part of the cost of a software
system. Maintenance includes enhancing the software, fixing errors identified as the
software is used, and modifying the software to work with new hardware and software.
It is essential to document programs thoroughly with titles and comment statements
because this greatly simplifies the task of software maintenance.
As discussed earlier, good programming technique plays an essential part in success-
ful DSP application. A structured and well-documented approach to programming
should be initiated from the beginning. It is important to develop an overall specifica-
tion for signal processing tasks prior to writing any program. The specification includes
the basic algorithm/task description, memory requirements, constraints on the program
size, execution time, etc. Specification review is an important component of the software
development process. A thoroughly reviewed specification can catch mistakes before
code is written and reduce potential code rework risk at system integration stage. The
potential use of subroutines for repetitive processes should also be noted. A flow
diagram will be a very helpful design tool to adopt at this stage. Program and data
blocks should be allocated to specific tasks that optimize data access time and address-
ing functions.
A software simulator or a hardware platform can be used for testing DSP code.
Software simulators run on a host computer to mimic the behavior of a DSP chip. The
DSP SYSTEM DESIGN 17
simulator is able to show memory contents, all the internal registers, I/O, etc., and the
effect on these after each instruction is performed. Input/output operations are simu-
lated using disk files, which require some format conversion. This approach reduces the
development process for software design only. Full real-time emulators are normally
used when the software is to be tested on prototype target hardware.
Writing and testing DSP code is a highly iterative process. With the use of a simulator
or an evaluation board, code may be tested regularly as it is written. Writing code in
modules or sections can help this process, as each module can be tested individually,
with a greater chance of the whole system working at the system integration stage.
There are two commonly used methods in developing software for DSP devices: an
assembly program or a C=C÷÷ program. Assembly language is one step removed from
the machine code actually used by the processor. Programming in assembly language
gives the engineers full control of processor functions, thus resulting in the most efficient
program for mapping the algorithm by hand. However, this is a very time-consuming
and laborious task, especially for today's highly paralleled DSP architectures. A C
program is easier for software upgrades and maintenance. However, the machine code
generated by a C compiler is inefficient in both processing speed and memory usage.
Recently, DSP manufactures have improved C compiler efficiency dramatically.
Often the ideal solution is to work with a mixture of C and assembly code. The overall
program is controlled by C code and the run-time critical loops are written in assembly
language. In a mixed programming environment, an assembly routine may be either
called as a function, or in-line coded into the C program. A library of hand-optimized
functions may be built up and brought into the code when required. The fundamentals
of C language for DSP applications will be introduced in Appendix C, while the
assembly programming for the TMS320C55x will be discussed in Chapter 2. Mixed C
and assembly programming will be introduced in Chapter 3. Alternatively, there are
many high-level system design tools that can automatically generate an implementation
in software, such as C and assembly language.
1.4.4 High-Level Software Development Tools
Software tools are computer programs that have been written to perform specific
operations. Most DSP operations can be categorized as being either analysis tasks
or filtering tasks. Signal analysis deals with the measurement of signal properties.
MATLAB is a powerful environment for signal analysis and visualization, which are
critical components in understanding and developing a DSP system. Signal filtering,
such as removal of unwanted background noise and interference, is usually a time-
domain operation. C programming is an efficient tool for performing signal filtering
and is portable over different DSP platforms.
In general, there are two different types of data files: binary files and ASCII (text)
files. A binary file contains data stored in a memory-efficient binary format, whereas an
ASCII file contains information stored in ASCII characters. A binary file may be
viewed as a sequence of characters, each addressable as an offset from the first position
in the file. The system does not add any special characters to the data except null
characters appended at the end of the file. Binary files are preferable for data that is
going to be generated and used by application programs. ASCII files are necessary if the
18 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
C program
(Source)
C compiler
Machine
code
(Object)
Linker/loader Execution
Program
output
Libraries Data
Figure 1.10 Program compilation, linking, and execution
data is to be shared by programs using different languages and different computer
platforms, especially for data transfer over computer networks. In addition, an ASCII
file can be generated using a word processor program or an editor.
MATLAB is an interactive, technical computing environment for scientific and
engineering numerical analysis, computation, and visualization. Its strength lies in the
fact that complex numerical problems can be solved easily in a fraction of the time
required with a programming language such as C. By using its relatively simple pro-
gramming capability, MATLAB can be easily extended to create new functions, and is
further enhanced by numerous toolboxes such as the Signal Processing Toolbox.
MATLAB is available on most commonly used computers such as PCs, workstations,
Macintosh, and others. The version we use in this book is based on MATLAB for
Windows, version 5.1. The brief introduction of using MATLAB for DSP is given in
Appendix B.
The purpose of a programming language is to solve a problem involving the manip-
ulation of information. The purpose of a DSP program is to manipulate signals in order
to solve a specific signal-processing problem. High-level languages are computer lan-
guages that have English-like commands and instructions. They include languages such
as C=C÷÷, FORTRAN, Basic, and Pascal. High-level language programs are usually
portable, so they can be recompiled and run on many different computers. Although
C is categorized as a high-level language, it also allows access to low-level routines.
In addition, a C compiler is available for most modern DSP devices such as the
TMS320C55x. Thus C programming is the most commonly used high-level language
for DSP applications.
C has become the language of choice for many DSP software development engineers
not only because it has powerful commands and data structures, but also because it can
easily be ported on different DSP platforms and devices. The processes of compilation,
linking/loading, and execution are outlined in Figure 1.10. A C compiler translates a
high-level C program into machine language that can be executed by the computer. C
compilers are available for a wide range of computer platforms and DSP chips, thus
making the C program the most portable software for DSP applications. Many C
programming environments include debugger programs, which are useful in identifying
errors in a source program. Debugger programs allow us to see values stored in
variables at different points in a program, and to step through the program line by line.
1.5 Experiments Using Code Composer Studio
The code composer studio (CCS) is a useful utility that allows users to create, edit,
build, debug, and analyze DSP programs. The CCS development environment supports
EXPERIMENTS USING CODE COMPOSER STUDIO 19
several Texas Instruments DSP processors, including the TMS320C55x. For building
applications, the CCS provides a project manager to handle the programming tasks.
For debugging purposes, it provides breakpoint, variable watch, memory/register/stack
viewing, probe point to stream data to and from the target, graphical analysis, execution
profiling, and the capability to display mixed disassembled and C instructions. One
important feature of the CCS is its ability to create and manage large projects from a
graphic-user-interface environment. In this section, we will use a simple sinewave
example to introduce the basic built-in editing features, major CCS components, and
the use of the C55x development tools. We will also demonstrate simple approaches to
software development and debugging process using the TMS320C55x simulator. The
CCS version 1.8 was used in this book.
Installation of the CCS on a PC or a workstation is detailed in the Code Composer
Studio Quick Start Guide [8]. If the C55x simulator has not been installed, use the
CCS setup program to configure and set up the TMS320C55x simulator. We can start
the CCS setup utility, either from the Windows start menu, or by clicking the Code
Composer Studio Setup icon. When the setup dialogue box is displayed as shown in
Figure 1.11(a), follow these steps to set up the simulator:
± Choose Install a Device Driver and select the C55x simulator device driver,
tisimc55.dvr for the TMS320C55x simulator. The C55x simulator will appear
in the middle window named as Available Board/Simulator Types if the installation
is successful, as shown in Figure 1.11(b).
± Drag the C55x simulator from Available Board/Simulator Types window to the
System Configuration window and save the change. When the system configuration
is completed, the window label will be changed to Available Processor Types as
shown in Figure 1.11(c).
1.5.1 Experiment 1A ± Using the CCS and the TMS320C55x Simulator
This experiment introduces the basic features to build a project with the CCS. The
purposes of the experiment are to:
(a) create projects,
(b) create source files,
(c) create linker command file for mapping the program to DSP memory space,
(d) set paths for C compiler and linker to search include files and libraries, and
(e) build and load program for simulation.
Let us begin with the simple sinewave example to get familiar with the TMS320C55x
simulator. In this book, we assume all the experiment files are stored on a disk in the
computer's A drive to make them portable for users, especially for students who may
share the laboratory equipment.
20 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
(a)
(b)
(c)
Figure 1.11 CCS setup dialogue boxes: (a) install the C55x simulator driver, (b) drag the C55x
simulator to system configuration window, and (c) save the configuration
The best way to learn a new software tool is by using it. This experiment is partitioned
into following six steps:
1. Start the CCS and simulator:
± Invoke the CCS from the Start menu or by clicking on the Code Composer Studio
icon on the PC. The CCS with the C55x simulator will appear on the computer
screen as shown in Figure 1.12.
2. Create a project for the CCS:
± Choose Project÷New to create a new project file and save it as exp1 to
A:\Experiment1. The CCS uses the project to operate its built-in utilities
to create a full build application.
3. Create a C program file using the CCS:
± Choose File÷New to create a new file, then type in the example C code listed
in Table 1.2, and save it as exp1.c to A:\Experiment1. This example reads
EXPERIMENTS USING CODE COMPOSER STUDIO 21
Figure 1.12 CCS integrated development environment
Table 1.2 List of sinewave example code, exp1.c
#define BUF_SIZE 40
const int sineTable[BUF_SIZE]=
{0x0000,0x000f,0x001e,0x002d,0x003a,0x0046,0x0050,0x0059,
0x005f,0x0062,0x0063,0x0062,0x005f,0x0059,0x0050,0x0046,
0x003a,0x002d,0x001e,0x000f,0x0000,0xfff1,0xffe2,0xffd3,
0xffc6,0xffba,0xffb0,0xffa7,0xffa1,0xff9e,0xff9d,0xff9e,
0xffa1,0xffa7,0xffb0,0xffba,0xffc6,0xffd3,0xffe2,0xfff1};
int in_buffer[BUF_SIZE];
int out_buffer[BUF_SIZE];
int Gain;
void main()
{
int i,j;
Gain = 0x20;
22 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
Table 1.2 (continued )
while (1)
{ /* <- set profile point on this line */
for (i = BUF_SIZE÷1; i >= 0; i÷÷)
{
j = BUF_SIZE÷1÷i;
out_buffer[j]= 0;
in_buffer[j]= 0;
}
for (i = BUF_SIZE÷1; i >= 0; i÷÷)
{
j = BUF_SIZE÷1÷i;
in_buffer[i]= sineTable[i]; /* <- set breakpoint */
in_buffer[i]= 0÷in_buffer[i];
out_buffer[j]= Gain*in_buffer[i];
}
} /* <- set probe and profile points on this line */
}
pre-calculated sinewave values from a table, negates, and stores the values in a
reversed order to an output buffer. Note that the program exp1.c is included in
the experimental software package.
However, it is recommended that we create this program with the editor to get
familiar with the CCS editing functions.
4. Create a linker command file for the simulator:
± Choose File÷New to create another new file and type in the linker
command file listed in Table 1.3 (or copy the file exp1.cmd from the experi-
mental software package). Save this file as exp1.cmd to A:\Experiment1.
Linker uses a command file to map different program segments into a pre-
partitioned system memory space. A detailed description on how to define and
use the linker command file will be presented in Chapter 2.
5. Setting up the project:
± After exp1.c and exp1.cmd are created, add them to the project by
choosing Project÷Add Files, and select files exp1.c and exp1.cmd from
A:\Experiment1.
± Before building a project, the search paths should be set up for the C
compiler, assembler, and linker. To set up options for the C compiler,
assembler, and linker, choose Project÷Options. The paths for the C55x
tools should be set up during the CCS installation process. We will need to add
search paths in order to include files and libraries that are not included in the
C55x tools directories, such as the libraries and included files we have created in
EXPERIMENTS USING CODE COMPOSER STUDIO 23
Table 1.3 Linker command file
/* Specify the system memory map */
MEMORY
{
RAM (RWIX): origin = 000100h, length = 01feffh /* Data memory */
RAM2 (RWIX): origin = 040100h, length = 040000h /* Program memory */
ROM (RIX) : origin = 020100h, length = 020000h /* Program memory */
VECS (RIX) : origin = 0ffff00h, length = 00100h /* Reset vector */
}
/* Specify the sections allocation into memory */
SECTIONS
{
vectors >VECS /* Interrupt vector table */
.text >ROM /* Code */
.switch >RAM /* Switch table information */
.const >RAM /* Constant data */
.cinit >RAM2 /* Initialization tables */
.data >RAM /* Initialized data */
.bss >RAM /* Global & static variables */
.sysmem >RAM /* Dynamic memory allocation area */
.stack >RAM /* Primary system stack */
}
± the working directory. For programs written in C language, it requires using the
run-time support library, rts55.lib for DSP system initialization. This can be
done by selecting Libraries under Category in the Linker dialogue box, and enter
the C55x run-time support library, rts55.lib. We can also specify different
directories to store the output executable file and map file. Figure 1.13 shows an
example of how to set the search paths for compiler, assembler, or linker.
6. Build and run the program:
± Once all the options are set, use Project÷Rebuild All command to build the
project. If there are no errors, the CCS will generate the executable output
file, exp1.out. Before we can run the program, we need to load the executable
output file to the simulator from File÷Load Program menu. Pick the file
exp1.out in A:\Experiment1 and open it.
± Execute this program by choosing Debug÷Run. DSP status at the bottom
left-hand corner of the simulator will be changed from DSP HALTED to DSP
RUNNING. The simulation process can be stopped withthe Debug÷Halt
command. We can continue the program by reissuing the run command or
exiting the simulator by choosing File÷Exit menu.
24 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
Figure 1.13 Setup search paths for C compiler, assembler, or linker
1.5.2 Experiment 1B ± Debugging Program on the CCS
The CCS has extended traditional DSP code generation tools by integrating a set of
editing, emulating, debugging, and analyzing capabilities in one entity. In this section of
the experiment, we will introduce some DSP program building steps and software
debugging capabilities including:
(a) the CCS standard tools,
(b) the advanced editing features,
(c) the CCS project environment, and
(d) the CCS debugging settings.
For a more detailed description of the CCS features and sophisticated configuration
settings, please refer to Code Composer Studio User's Guide [7].
Like most editors, the standard tool bar in Figure 1.12 allows users to create and open
files, cut, copy, and paste texts within and between files. It also has undo and re-do
capabilities to aid file editing. Finding or replacing texts can be done within one file or in
different files. The CCS built-in context-sensitive help menu is also located in the
standard toolbar menu. More advanced editing features are in the edit toolbar menu,
refer to Figure 1.12. It includes mark to, mark next, find match, and find next open
parenthesis capabilities for C programs. The features of out-indent and in-indent can be
used to move a selected block of text horizontally. There are four bookmarks that allow
users to create, remove, edit, and search bookmarks.
EXPERIMENTS USING CODE COMPOSER STUDIO 25
The project environment contains a C compiler, assembler, and linker for users to
build projects. The project toolbar menu (see Figure 1.12) gives users different choices
while working on projects. The compile only, incremental build, and build all functions
allow users to build the program more efficiently. Breakpoints permit users to set stop
points in the program and halt the DSP whenever the program executes at those
breakpoint locations. Probe points are used to transfer data files in and out of pro-
grams. The profiler can be used to measure the execution time of the program. It
provides program execution information, which can be used to analyze and identify
critical run-time blocks of the program. Both the probe point and profile will be
discussed in detail in the next section.
The debug toolbar menu illustrated in Figure 1.12 contains several step operations:
single step, step into a function, step over a function, and step out from a function back
to its caller function. It can also perform the run-to-cursor operation, which is a very
convenient feature that allows users to step through the code. The next three hot
buttons in the debug tool bar are run, halt, and animate. They allow users to execute,
stop, and animate the program at anytime. The watch-windows are used to monitor
variable contents. DSP CPU registers, data memory, and stack viewing windows
provide additional information for debugging programs. More custom options are
available from the pull-down menus, such as graphing data directly from memory
locations.
When we are developing and testing programs, we often need to check the values of
variables during program execution. In this experiment, we will apply debugging
settings such as breakpoints, step commands, and watch-window to understand the
CCS. The experiment can be divided into the following four steps.
1. Add and remove breakpoints:
± Start with Project÷Open, select exp1 in the A:\Experiment1 directory.
Build and load the experiment exp1.out. Double-click on the C file exp1.c in
the project-viewing window to open it from the source folder.
± Adding and removing a breakpoint to a specific line is quite simple. To add a
breakpoint, move the cursor to the line where we want to set a breakpoint. The
command to enable a breakpoint can be given from the Toggle Breakpoint hot
button on the project toolbar or by clicking the right mouse button and choosing
toggle breakpoint. The function key <F9> is a shortcut key that also enables
breakpoints. Once a breakpoint is enabled, a red dot will appear on the left to
indicate where the breakpoint is set. The program will run up to that line without
exceeding it. To remove breakpoints, we can either toggle breakpoints one by one,
or we can select the Delete All tap from the debug tool bar to clear all the break-
points at once. Now put the cursor on the following line:
in_buffer[i] = sineTable[i]; /* <- set breakpoint */
and click the Toggle Breakpoint toolbar button (or press <F9>).
2. Set up viewing windows:
± On the standard tool menu bar click View÷CPU Registers÷CPU Registers
to open the CPU registers window. We can edit the contents of any
26 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
CPU register by double clicking on it. Right click on the CPU Register window
and select Allow Docking. We can now move and resize the window. Try
to change the temporary register T0 and accumulator AC0 to T0 = 0x1234 and
AC0 = 0x56789ABC.
± On the CCS menu bar click Tools÷Command Window to add the Command
Window. We can resize and dock it as in the previous step. The command
window will appear each time we rebuild the project.
± We can customize the CCS display and settings using the workspace
feature. To save a workspace, click File÷Workspace÷Save Workspace and give
the workspace a name. When we restart CCS, we can reload that workspace by
clicking File÷Workspace÷Load Workspace and select the proper workspace
filename.
± Click View÷Dis-Assembly on the menu bar to see the disassembly window.
Every time we reload an executable file, the disassembly window will appear
automatically.
3. Using the single step features:
± When using C programs, the C55x system uses a function called boot from the
run-time support library rts55.lib to initialize the system. After we load the
exp1.out, the program counter (PC) should be at the start of the boot function
and the assembly code, boot.asm, should be displayed in the disassembly
window. For a project starting with C programs, there must be a function called
main() from which the C functions logically begin to execute. We can issue the
command, Go Main, from the Debug menu to start the C program.
± After the Go Main command, the DSP will be halted at the location where the
function main() is. Hit the <F8> key or click the single step button on the debug
toolbar repeatedly and single-step through the program exp1.c, watching the
values of the CPU registers change. Move the cursor to a different location in the
code and try the run-to-cursor command (hold down the <Ctrl>and <F10>keys
simultaneously).
4. Resource monitoring:
± From View÷Watch Window, open the Watch Window area. At run time, this
area shows the values of watched variables. Right-click on the Watch Window
area and choose Insert New Expression from the pop-up list. Type the output
buffer name, out_buffer, into the expression box and click OK, expand the
out_buffer to view each individual element of the buffer.
± From View÷Memory, open a memory window and enter the starting address
of the in_buffer in data page to view the data in the input and output buffers.
Since global variables are defined globally, we can use the variable name as the
address for memory viewing.
± From View÷Graph÷Time/Frequency, open the Graphic Property dialogue. Set
the display parameters as shown in Figure 1.14. The CCS allows the user to plot
data directly from memory by specifying the memory location and its length.
± Set a breakpoint on the line of the following C statement:
EXPERIMENTS USING CODE COMPOSER STUDIO 27
Figure 1.14 Graphics display settings
in_buffer[i] = sineTable[i]; /* <- set breakpoint */
Start animation execution, and view CPU registers, in_buffer and out_buffer data
in both the watch-window and the memory window. Figure 1.15 shows one instant
snapshot of the animation. The yellow arrow represents the current program counter's
location, and the red dot shows where the breakpoint is set. The data and register values
in red color are the ones that have just been updated.
1.5.3 Experiment 1C ± File Input and Output
Probe point is a useful tool for algorithm development, such as simulating real-time
input and output operations. When a probe point is reached, the CCS can either read a
selected amount of data samples from a file on the host PC to DSP memory on the
target, or write processed data samples to the host PC. In the following experiment, we
will learn how to set up a probe point to transfer data from the example program to a
file on the host computer.
± Set the probe point at the end of the while {} loop at the line of the close bracket
as:
while(1)
{
... ...
} /* <- set probe point on this line */
± where the data in the output buffer is ready to be transferred out. Put the cursor on
the line and click Toggle Probe Point. A blue dot on the left indicates the probe point
is set (refer to Figure 1.15).
28 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
Figure 1.15 CCS screen snapshot of the Experiment 1B
± From File÷File I/O, open the file I/O dialogue box and select File Output tab.
From the Add File tab, enter exp1_out.dat as file name, then select Open. Using
the output variable name, out_buffer, as the address and 40 (BUF_SIZE) as the
length of the data block for transferring 40 data samples to the host computer
from the buffer every time the probe point is reached. Now select Add Probe Point
tab to connect the probe point with the output file exp1_out.dat as shown in
Figure 1.16.
± Restart the program. After execution, we can view the data file exp1_out.dat
using the built-in editor by issue File÷Open command. If we want to view
or edit the data file using other editors/viewers, we need to exit the CCS or
disconnect the file from the File I/O.
An example data file is shown in Table 1.4. The first line contains the header
information in TI Hexadecimal format, which uses the syntax illustrated in Figure 1.17.
For the example given in Table 1.4, the data stored is in hexadecimal format with the
address of out_buffer at 0xa8 on data page and each block containing 40 (0x28) data
values. If we want to use probe to connect an input data file to the program, we will
need to use the same hex format to include a header in the input data file.
1.5.4 Experiment 1D ± Code Efficiency Analysis
The profiler can be used to measure the system execution status of specific segments of
the code. This feature gives users an immediate result about the program's performance.
EXPERIMENTS USING CODE COMPOSER STUDIO 29
(a)
(b)
Figure 1.16 Connect probe point to a file: (a) set up probe point address and length, and
(b) connect probe point with a file
Table 1.4 Data file saved by CCS
1651 1 a8 1 28
0x01E0
0x03C0
0x05A0
0x0740
0x08C0
0x0A00
0x0B20
0x0BE0
. . .
30 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
Figure 1.17 CCS File header format
It is a very useful tool for analyzing and optimizing DSP code for large complex
projects. In the following experiment, we use the profiling features of the CCS to obtain
statistics about code execution time.
± Open the project exp1 and load the file exp1.out. Open the source file exp1.c
and identify the line numbers on the source code where we like to set profile
marks. For a demonstration purpose, we will profile the entire code within the
while {}loop in the experiment. The profile points are set at line 32 and 46 as
shown below:
while (1)
{ /* <- set profile point here */
... ...
} /* <- set profile point here */
± From Profiler menu, select Start New Session to open the profile window. Click the
Create Profile Area hot button, and in the Manual Profile Area Creation dialogue
box (see Figure 1.18), enter the number for starting and ending lines. In the mean-
time, make sure the Source File Lines and the Generic type are selected. Finally,
click on the Ranges tab to switch the window that displays the range of the code
segments we just selected.
± The profiler is based on the system clock. We need to select Profile÷
± Enable Clock to enable the system clock before starting the profiler. This clock
counts instruction cycles. The clock setting can also be adjusted. Since the C55x
simulator does not have the connection to real hardware, the profiler of the simu-
lator can only display CPU cycles in the count field (refer to the example given in
Figure 1.18). More information can be obtained by using real DSP hardware such as
the C55x EVMs.
± Run the program and record the cycle counts shown in the profile status
window.
EXPERIMENTS USING CODE COMPOSER STUDIO 31
Figure 1.18 Profile window displaying DSP run-time status
Table 1.5 Gain control GEL function
Menuitem "Gain Control"
slider Gain(1, 0x20, 1, 1, gainParam)
{
Gain = gainParam;
}
Figure 1.19 GEL slide bar
1.5.5 Experiment 1E ± General Extension Language
The CCS uses General Extension Language (GEL) to extend its functions. GEL is a
useful tool for automated testing and design of workspace customization. The Code
Composer Studio User's Guide [7] provides a detailed description of GEL functions. In
this experiment, we will use a simple example to introduce it.
Create a file called Gain.gel and type the simple GEL code listed in Table
1.5. From the CCS, load this GEL file from File÷Load GEL and bring the Gain
32 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
control slide bar shown in Figure 1.19 out from GEL÷Gain Control. While
animating the program using the CCS, we can change the gain by moving the slider up
and down.
References
[1] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, Englewood Cliffs, NJ:
Prentice-Hall, 1989.
[2] S. J. Orfanidis, Introduction to Signal Processing, Englewood Cliffs, NJ: Prentice-Hall, 1996.
[3] J. G. Proakis and D. G. Manolakis, Digital Signal Processing ± Principles, Algorithms, and Applica-
tions, 3rd Ed., Englewood Cliffs, NJ: Prentice-Hall, 1996.
[4] A Bateman and W. Yates, Digital Signal Processing Design, New York: Computer Science Press,
1989.
[5] S. M. Kuo and D. R. Morgan, Active Noise Control Systems ± Algorithms and DSP Implementa-
tions, New York: Wiley, 1996.
[6] J. H. McClellan, R. W. Schafer, and M. A. Yoder, DSP First: A Multimedia Approach, 2nd Ed.,
Englewood Cliffs, NJ: Prentice-Hall, 1998.
[7] Texas Instruments, Inc., Code Composer Studio User's Guide, Literature no. SPRU32, 1999.
[8] Texas Instruments, Inc., Code Composer Studio Quick Start Guide, Literature no. SPRU368A, 1999.
Exercises
Part A
1. Given an analog audio signal with frequencies up to 10 kHz.
(a) What is the minimum required sampling frequency that allows a perfect reconstruction of
the signal from its samples?
(b) What will happen if a sampling frequency of 8 kHz is used?
(c) What will happen if the sampling frequency is 50 kHz?
(d) When sampled at 50 kHz, if only taking every other samples (this is a decimation by 2),
what is the frequency of the new signal? Is this causing aliasing?
2. Refer to Example 1.1, assuming that we have to store 50 ms (milliseconds, 1 ms = 10
÷3
seconds)
of digitized signals. How many samples are needed for (a) narrowband telecommunication
systems with f
s
= 8 kHz, (b) wideband telecommunication systems with f
s
= 16 kHz, (c) audio
CDs with f
s
= 44:1 kHz, and (d) professional audio systems with f
s
= 48 kHz.
3. Given a discrete time sinusoidal signal of x(n) = 5 sin(np=100) V.
(a) Find its peak-to-peak range?
(b) What is the quantization resolution of an 8-bit ADC for this signal?
(c) In order to obtain the quantization resolution of below 1 mV, how many bits are required
in the ADC?
EXERCISES 33
Part B
4. From the Option menu, set the CCS for automatically loading the program after the
project has been built.
5. To reduce the number of mouse click, many pull-down menu items have been mapped
to the hot buttons for the standard, advanced edit, project management, and debug
tools bar. There are still some functions; however, do not associate with any hot
buttons. Using the Option menu to create shortcut keys for the following menu items:
(a) map Go Main in the debug menu to Alt÷M (Alt key and M key),
(b) map Reset in the debug menu to Alt÷R,
(c) map Restart in the debug menu to Alt÷S, and
(d) map File reload in the file menu to Ctrl÷R.
6. After having loaded a program into the simulator and enabled Source/ASM mixed
display mode from View÷Mixed Source/ASM, what is showing in the CCS source display
window besides the C source code?
7. How to change the format of displayed data in the watch-window to hex, long, and
floating-point format from integer format?
8. What does File÷Workspace do? Try the save and reload workspace commands.
9. Besides using file I/O with the probe point, data values in a block of memory
space can also be stored to a file. Try the File÷Data÷Store and File÷Data÷Load
commands.
10. Use Edit÷Memory command we can manipulate (edit, copy, and fill) system memory:
(a) open memory window to view out_buffer,
(b) fill out_buffer with data 0x5555, and
(c) copy the constant sineTable[]to out_buffer.
11. Using CCS context-sensitive on-line help menu to find the TMS320C55x CUP diagram,
and name all the buses and processing units.
34 INTRODUCTION TO REAL-TIME DIGITAL SIGNAL PROCESSING
2
Introduction to TMS320C55x
Digital Signal Processor
Digital signal processors with architecture and instructions specifically designed for
DSP applications have been launched by Texas Instruments, Motorola, Lucent Tech-
nologies, Analog Devices, and many other companies. DSP processors are widely used
in areas such as communications, speech processing, image processing, biomedical
devices and equipment, power electronics, automotive, industrial electronics, digital
instruments, consumer electronics, multimedia systems, and home appliances.
To efficiently design and implement DSP systems, we must have a solid knowledge of
DSP algorithms as well as a basic concept of processor architecture. In this chapter, we
will introduce the architecture and assembly programming of the Texas Instruments
TMS320C55x fixed-point processor.
2.1 Introduction
Wireless communications, telecommunications, medical, and multimedia applications
are developing rapidly. Increasingly traditional analog devices are being replaced
with digital systems. The fast growth of DSP applications is not a surprise when
considering the commercial advantages of DSP in terms of the potentially fast time to
market, flexibility for upgrades to new technologies and standards, and low design
cost offered by various DSP devices. The rising demand from the digital handheld
devices in the consumer market to the digital networks and communication infrastruc-
tures coupled with the emerging internet applications are the driving forces for DSP
applications.
In 1982, Texas Instruments introduced its first general-purpose fixed-point DSP
device, the TMS32010, to the consumer market. Since then, the TMS320 family
has extended into two major classes: the fixed-point and floating-point processors.
The TMS320 fixed-point family consists of C1x, C2x, C5x, C2xx, C54x, C55x, C62x,
and C64x. The TMS320 floating-point family includes C3x, C4x, and C67x. Each
generation of the TMS320 series has a unique central processing unit (CPU) with
a variety of memory and peripheral configurations. In this book, we chose the
TMS320C55x as an example for real-time DSP implementations, applications, and
experiments.
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
The C55x processor is designed for low power consumption, optimum performance,
and high code density. Its dual multiply±accumulate (MAC) architecture provides twice
the cycle efficiency computing vector products ± the fundamental operation of digital
signal processing, and its scaleable instruction length significantly improves the code
density. In addition, the C55x is source code compatible with the C54x. This greatly
reduces the migration cost from the popular C54x based systems to the C55x systems.
Some essential features of the C55x device are listed below:
. Upward source-code compatible with all TMS320C54x devices.
. 64-byte instruction buffer queue that works as a program cache and efficiently
implements block repeat operations.
. Two 17-bit by 17-bit MAC units can execute dual multiply-and-accumulate oper-
ations in a single cycle.
. A 40-bit arithmetic and logic unit (ALU) performs high precision arithmetic and
logic operations with an additional 16-bit ALU performing simple arithmetic
operations parallel to the main ALU.
. Four 40-bit accumulators for storing computational results in order to reduce
memory access.
. Eight extended auxiliary registers for data addressing plus four temporary data
registers to ease data processing requirements.
. Circular addressing mode supports up to five circular buffers.
. Single-instruction repeat and block repeat operations of program for supporting
zero-overhead looping.
Detailed information about the TMS320C55x can be found in the manufacturer's
manuals listed in references [1±6].
2.2 TMS320C55x Architecture
The C55x CPU consists of four processing units: an instruction buffer unit (IU), a
program flow unit (PU), an address-data flow unit (AU), and a data computation unit
(DU). These units are connected to 12 different address and data buses as shown in
Figure 2.1.
2.2.1 TMS320C55x Architecture Overview
Instruction buffer unit (IU): This unit fetches instructions from the memory into the
CPU. The C55x is designed for optimum execution time and code density. The instruc-
tion set of the C55x varies in length. Simple instructions are encoded using eight bits
36 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
BB

CB DB 32 bits
Data
computation
unit
(DU)
Program
flow unit
(PU)
Address data
flow unit
(AU)
C55x CPU
Instruction
buffer unit
(IU)
Two 24-bit data-write address buses (EAB, FAB)
24-bit program-read address bus (PAB)
32-bit program-read data bus (PB)
Three 16-bit data-read data buses (BB, CB, DB)
Three 24-bit data-read address buses (BAB, CAB, DAB)
CB DB
Two 16-bit data-write data buses (EB, FB)
Figure 2.1 Block diagram of TMS320C55x CPU
32 (4-byte opcode fetch)
IU
(1-6 bytes
opcode)
48
Instruction
buffer
queue
(64 bytes)
Instruction
decoder
PU
AU
DU
Program-read data bus (PB)
Figure 2.2 Simplified block diagram of the C55x instruction buffer unit
(one byte), while more complicated instructions may contain as many as 48 bits (six
bytes). For each clock cycle, the IU can fetch four bytes of program code via its 32-bit
program-read data bus. At the same time, the IU can decode up to six bytes of program.
After four program bytes are fetched, the IU places them into the 64-byte instruction
buffer. At the same time, the decoding logic decodes an instruction of one to six bytes
previously placed in the instruction decoder as shown in Figure 2.2. The decoded
instruction is passed to the PU, the AU, or the DU.
The IU improves the efficiency of the program execution by maintaining a constant
stream of instruction flow between the four units within the CPU. If the IU is able to
TMS320C55X ARCHITECTURE 37
hold a segment of the code within a loop, the program execution can be repeated many
times without fetching additional code. Such a capability not only improves the loop
execution time, but also saves the power consumption by reducing program accesses
from the memory. Another advantage is that the instruction buffer can hold multiple
instructions that are used in conjunction with conditional program flow control. This
can minimize the overhead caused by program flow discontinuities such as conditional
calls and branches.
Program flow unit (PU): This unit controls DSP program execution flow. As illus-
trated in Figure 2.3, the PU consists of a program counter (PC), four status registers, a
program address generator, and a pipeline protection unit. The PC tracks the C55x
program execution every clock cycle. The program address generator produces a 24-bit
address that covers 16 Mbytes of program space. Since most instructions will be exe-
cuted sequentially, the C55x utilizes pipeline structure to improve its execution effi-
ciency. However, instructions such as branches, call, return, conditional execution, and
interrupt will cause a non-sequential program address switch. The PU uses a dedicated
pipeline protection unit to prevent program flow from any pipeline vulnerabilities
caused by a non-sequential execution.
Address-data flow unit (AU): The address-data flow unit serves as the data access
manager for the data read and data write buses. The block diagram illustrated in Figure
2.4 shows that the AU generates the data-space addresses for data read and data write.
It also shows that the AU consists of eight 23-bit extended auxiliary registers (XAR0±
XAR7), four 16-bit temporary registers (T0±T3), a 23-bit extended coefficient data
pointer (XCDP), and a 23-bit extended stack pointer (XSP). It has an additional 16-
bit ALU that can be used for simple arithmetic operations. The temporary registers may
be utilized to expand compiler efficiency by minimizing the need for memory access. The
AU allows two address registers and a coefficient pointer to be used together for
processing dual-data and one coefficient in a single clock cycle. The AU also supports
up to five circular buffers, which will be discussed later.
Data computation unit (DU): The DU handles data processing for most C55x
applications. As illustrated in Figure 2.5, the DU consists of a pair of MAC units, a
40-bit ALU, four 40-bit accumulators (AC0, AC1, AC2, and AC3), a barrel shifter,
rounding and saturation control logic. There are three data-read data buses that
allow two data paths and a coefficient path to be connected to the dual-MAC units
simultaneously. In a single cycle, each MAC unit can perform a 17-bit multiplication
24-bit
Program-read address bus (PAB)
Program counter (PC)
Status registers
(ST0, ST1, ST2, ST3)
Address generator
Pipeline protection unit
PU
Figure 2.3 Simplified block diagram of the C55x program flow unit
38 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
FB
EB
FAB
EAB
BAB
CAB
DAB
CB

DB
D
A
T
A
M
E
M
O
R
Y
S
P
A
C
E
XAR0
XAR1
XAR2
XAR3
XAR4
XAR5
XAR6
XAR7
XCDP
XSP
16-bit
ALU
T0
T1
T2
T3
16-bit 23-bit
AU
Data
address
generator
unit
(24-bit)
Figure 2.4 Simplified block diagram of the C55x address-data flow unit
DB
CB
BB
FB
16-bit
EB
16-bit
DU
ALU
(40-bit)
Barrel
Shifter
Overflow
&
Saturation
AC0
AC1
AC2
AC3
MAC
MAC
16-bit
16-bit
16-bit
Figure 2.5 Simplified block diagram of the C55x data computation unit
and a 40-bit addition or subtraction operation with a saturation option. The ALU can
perform 40-bit arithmetic, logic, rounding, and saturation operations using the four
accumulators. It can also be used to achieve two 16-bit arithmetic operations in both the
upper and lower portions of an accumulator at the same time. The ALU can accept
immediate values from the IU as data and communicate with other AU and PU
registers. The barrel shifter may be used to perform a data shift in the range of 2
À32
(shift right 32-bit) to 2
31
(shift left 31-bit).
2.2.2 TMS320C55x Buses
As illustrated in Figure 2.1, the TMS320C55x has one 32-bit program data bus, five 16-
bit data buses, and six 24-bit address buses. The program buses include a 32-bit
program-read data bus (PB) and a 24-bit program-read address bus (PAB). The PAB
carries the program memory address to read the code from the program space. The unit
of program address is in bytes. Thus the addressable program space is in the range of
TMS320C55X ARCHITECTURE 39
0x000000±0xFFFFFF (the prefix 0x indicates the following number is in hexadecimal
format). The PB transfers four bytes of program code to the IU each clock cycle.
The data buses consist of three 16-bit data-read data buses (BB, CB, and DB) and
three 24-bit data-read addresses buses (BAB, CAB, and DAB). This architecture sup-
ports three simultaneous data reads from data memory or I/O space. The C bus and D
buses (CB and DB) can send data to the PU, AU, and DU; while the B bus (BB) can
only work with the DU. The primary function of the BB is to connect memory to a dual-
MAC; so some specific operations can access all three data buses, such as fetching two
data and one coefficient. The data-write operations are carried out using two 16-bit
data-write data buses (EB and FB) and two 24-bit data-write address buses (EAB and
FAB). For a single 16-bit data write, only the EB is used. A 32-bit data write will use
both the EB and FB in one cycle. The data-write address buses (EAB and FAB)
have the same 24-bit addressing range. Since the data access uses a word unit (2-byte),
the data memory space becomes 23-bit word addressable from address 0x000000 to
0x7FFFFF.
The C55x architecture is built around these 12 buses. The program buses carry the
instruction code and immediate operands from program memory, while the data buses
connect various units. This architecture maximizes the processing power by maintaining
separate memory bus structures for full-speed execution.
2.2.3 TMS320C55x Memory Map
The C55x uses a unified program, data, and I/O memory configurations. All 16 Mbytes
of memory are available as program or data space. The program space is used for
instructions and the data space is used for general-purpose storage and CPU memory
mapped registers. The I/O space is separated from the program/data space, and is used
for duplex communication with peripherals. When the CPU fetches instructions from
the program space, the C55x address generator uses the 24-bit program-read address
bus. The program code is stored in byte units. When the CPU accesses data space, the
C55x address generator masks the least-significant-bit (LSB) of the data address since
data stored in memory is in word units. The 16 Mbytes memory map is shown in Figure
2.6. Data space is divided into 128 data pages (0±127). Each page has 64 K words. The
memory block from address 0 to 0x5F in page 0 is reserved for memory mapped
registers (MMRs).
2.3 Software Development Tools
The manufacturers of DSP processors typically provide a set of software tools for the
user to develop efficient DSP software. The basic software tools include an assembler,
linker, C compiler, and simulator. As discussed in Section 1.4, DSP programs can be
written in either C or assembly language. Developing C programs for DSP applications
requires less time and effort than those applications using assembly programs. However,
the run-time efficiency and the program code density of the C programs are generally
worse than those of the assembly programs. In practice, high-level language tools such
40 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
MMRs 00 0000-00 005F 00 0000-00 00BF Reserved
00 0060
00 FFFF
00 00C0
01 FFFF
01 0000
01 FFFF
02 0000
03 FFFF
02 0000
02 FFFF
04 0000
05 FFFF
7F 0000
7F FFFF
FE 0000
FF FFFF
Page 0

Page 1

Page 2

Page 127

Program space addresses
byte in Hexadecimal
C55x memory
program/data space
Data space addresses
word in Hexadecimal
Figure 2.6 TMS320C55x program space and data space memory map
as MATLAB and C are used in early development stages to verify and analyze the
functionality of the algorithms. Due to real-time constraints and/or memory limitations,
part (or all) of the C functions have to be replaced with assembly programs.
In order to execute the designed DSP algorithms on the target system, the C or
assembly programs must first be translated into binary machine code and then linked
together to form an executable code for the target DSP hardware. This code conversion
process is carried out using the software development tools illustrated in Figure 2.7.
The TMS320C55x software development tools include a C compiler, an assembler,
a linker, an archiver, a hex conversion utility, a cross-reference utility, and an absolute
lister. The debugging tools can either be a simulator or an emulator. The C55x C
compiler generates assembly code from the C source files. The assembler translates
assembly source files; either hand-coded by the engineers or generated by the C com-
piler, into machine language object files. The assembly tools use the common object file
format (COFF) to facilitate modular programming. Using COFF allows the program-
mer to define the system's memory map at link time. This maximizes performance by
enabling the programmer to link the code and data objects into specific memory
locations. The archiver allows users to collect a group of files into a single archived
file. The linker combines object files and libraries into a single executable COFF object
module. The hex conversion utility converts a COFF object file into a format that can
be downloaded to an EPROM programmer.
In this section, we will briefly describe the C compiler, assembler, and linker. A full
description of these tools can be found in the user's guides [2,3].
SOFTWARE DEVELOPMENT TOOLS 41
C
source files
C compiler
Assembler
Assembly
source files
COFF
object files
Linker
Run-time
support
libraries
COFF
executable
file
Library-build
utility
Archiver
Library of
object files
Archiver
Macro
library
Macro
source files
TMS320C55x
target
Absolute
lister
X-reference
lister
Hex
converter
EPROM
programmer
Debugger
Figure 2.7 TMS320C55x software development flow and tools
2.3.1 C Compiler
As mentioned in Chapter 1, C language is the most popular high-level tool for evaluating
DSP algorithms and developing real-time software for practical applications. The
TMS320C55x C compiler translates the C source code into the TMS320C55x assembly
source code first. The assembly code is then given to the assembler for generating machine
code. The C compiler can generate either a mnemonic assembly code or algebraic
assembly code. Table 2.1 gives an example of the mnemonic and algebraic assembly
code generated by the C55x compiler. In this book, we will introduce only the widely used
mnemonic assembly language. The C compiler package includes a shell program, code
optimizer, and C-to-ASM interlister. The shell program supports automatic compile,
assemble, and link modules. The optimizer improves run-time and code density efficiency
of the C source files. The C-to-ASM interlister inserts the original comments in C source
code into the compiler's output assembly code; so the user can view the corresponding
assembly instructions generated by the compiler for each C statement.
The C55x compiler supports American National Standards Institute (ANSI) C and its
run-time-support library. The run-time support library, rts55.lib, includes functions
to support string operation, memory allocation, data conversion, trigonometry, and
exponential manipulations. The CCS introduced in Section 1.5 has made using DSP
development tools (compiler, assembly, and linker) easier by providing default setting
42 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
Table 2.1 An example of C code and the C55x compiler generated assembly code
Code Mnemonic assembly code Algebraic assembly code
mov *SP(#0), AR2 AR2 ˆ *SP(#0)
add #_sineTable, AR2 AR2 ˆ AR2 ‡ #_sineTable
in_buffer[i] ˆ sineTable[i]; mov *SP(#0), AR3 AR3 ˆ *SP(#0)
add #_in_buffer, AR3 AR3 ˆ AR3 ‡ #_in_buffer
mov *AR2, *AR3 *AR3 ˆ *AR2
parameters and prompting the options. It is still beneficial for the user to understand
how to use these tools individually, and set parameters and options from the command
line correctly.
We can invoke the C compiler from a PC or workstation shell by entering the
following command:
c155 [-options] [filenames] [-z[link_options] [object_files]]
The filenames can be one or more C program source files, assembly source files,
object files, or a combination of these files. If we do not supply an extension, the
compiler assumes the default extension as .c, .asm, or .obj. The -z option enables
the linker, while the -c option disables the linker. The link_options set up the way
the linker processes the object files at link time. The object_files are additional
objective files for the linker to add to the target file at link time. The compiler options
have the following categories:
1. The options that control the compiler shell, such as the -g option that generates
symbolic debug information for debugging code.
2. The options that control the parser, such as the -ps option that sets the strict ANSI
C mode for C.
3. The options that are C55x specific, such as the -ml option that sets the large
memory model.
4. The options that control the optimization, such as the -o0 option that sets the
register optimization.
5. The options that change the file naming conventions and specify the directories,
such as the -eo option that sets the default object file extension.
6. The options that control the assembler, such as the -al option that creates assem-
bly language listing files.
7. The options that control the linker, such as the -ar option that generates a re-
locatable output module.
SOFTWARE DEVELOPMENT TOOLS 43
There are a number of options in each of the above categories. Refer to the
TMS320C55x Optimizing C Compiler User's Guide [3] for detailed information on
how to use these options.
The options are preceded by a hyphen and are not case sensitive. All the single letter
options can be combined together, i.e., the options of -g, -k, and -s, are the same as
setting the compiler options as -gks. The two-letter operations can also be combined if
they have the same first letter. For example, setting -pl, -pk, and -pi three options are
the same as setting the options as -plki.
C language lacks specific DSP features, especially those of fixed-point data oper-
ations that are necessary for many DSP algorithms. To improve compiler efficiency for
real-time DSP applications, the C55x compiler provides a method to add in-line assem-
bly language routines directly into the C program. This allows the programmer to write
highly efficient assembly code for the time-critical sections of a program. Intrinsic is
another improvement for users to substitute DSP arithmetic operation with assembly
intrinsic operators. We will introduce more compiler features in Section 2.7 when we
present the mixing of C and assembly programs. In this chapter, we emphasize assembly
language programming.
2.3.2 Assembler
The assembler translates processor-specific assembly language source files (in ASCII
text) into binary COFF object files for specific DSP processors. Source files can contain
assembler directives, macro directives, and instructions. Assembler directives are used to
control various aspects of the assembly process such as the source file listing format,
data alignment, section content, etc. Binary object files contain separate blocks (called
sections) of code or data that can be loaded into memory space.
Assembler directives are used to control the assembly process and to enter data
into the program. Assembly directives can be used to initialize memory, define global
variables, set conditional assembly blocks, and reserve memory space for code and data.
Some of the most important C55x assembler directives are described below:
.BSS directive: The .bss directive reserves space in the uninitialized .bss section for
data variables. It is usually used to allocate data into RAM for run-time variables such
as I/O buffers. For example,
.bss xn_buffer, size_in_words
where the xn_buffer points to the first location of the reserved memory space, and the
size_in_words specifies the number of words to be reserved in the .bss section. If
we do not specify uninitialized data sections, the assembler will put all the uninitialized
data into the .bss section.
.DATA directive: The .data directive tells the assembler to begin assembling the
source code into the .data section, which usually contains data tables or pre-initialized
variables such as sinewave tables. The data sections are word addressable.
.SECT directive: The .sect directive defines a section and tells the assembler to
begin assembling source code or data into that section. It is often used to separate long
programs into logical partitions. It can separate the subroutines from the main pro-
gram, or separate constants that belong to different tasks. For example,
44 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
.sect "section_name"
assigns the code into the user defined memory section called section_name. Code
from different source files with the same section names are placed together.
.USECT directive: The .usect reserves space in an uninitialized section. It is similar
to the .bss directive. It allows the placement of data into user defined sections instead
of.bss sections. It is often used to separate large data sections into logical partitions,
such as separating the transmitter data variables from the receiver data variables. The
syntax of .usect directive is
symbol .usect "section_name", size_in_words
where symbol is the variable, or the starting address of a data array, which will be
placed into the section named section_name. In the latter case, the size_in_words
defines the number of words in the array.
.TEXT directive: The .text directive tells the assembler to begin assembling source
code into the .text section, which normally contains executable code. This is the
default section for program code. If we do not specify a program code section, the
assembler will put all the programs into the .text section.
The directives, .bss, .sect, .usect, and.text are used to define the memory
sections. The following directives are used to initialize constants.
.INT (.WORD) directive: The .int (or .word) directive places one or more 16-bit
integer values into consecutive words in the current section. This allows users to
initialize memory with constants. For example,
data1 .word 0x1234
data2 .int 1010111b
In these examples, data1 is initialized to the hexadecimal number 0x1234 (decimal
number 4660), while data2 is initialized to the binary number of 1010111b (decimal
87). The suffix `b' indicates the data 1010111 is in binary format.
.SET (.EQU) directive: The .set (or .equ) directive assigns values to symbols. This
type of symbol is known as an assembly-time constant. It can then be used in source
statements in the same manner as a numeric constant. The .set directive has the
form:
symbol .set value
where the symbol must appear in the first column. This example equates the constant
value to the symbol. The symbolic name used in the program will be replaced with the
constant by the assembler during assembly time, thus allowing programmers to write
more readable programs. The .set and .equ directives can be used interchangeably,
and do not produce object code.
The assembler is used to convert assembly language source code to COFF format
object files for the C55x processor. The following command invokes the C55x mne-
monic assembler:
masm55 [input_file [object_file [list_file]]] [-options]
The input_file is the name of the assembly source program. If no extension is
supplied, the assembler assumes that the input_file has the default extension
SOFTWARE DEVELOPMENT TOOLS 45
.asm. The object_file is the name of the object file that the assembler creates. The
assembler uses the source file's name with the default extension .obj for the object
file unless specified otherwise. The list_file is the name of the list file that the
assembler creates. The assembler will use the source file's name and .lst as the default
extension for the list file. The assembler will not generate list files unless the option -l
is set.
The options identify the assembler options. Some commonly used assembler
options are:
. The -l option tells the assembler to create a listing file showing where the program
and the variables are allocated.
. The -s option puts all symbols defined in the source code into the symbol table so
the debugger may access them.
. The -c option makes the case insignificant in symbolic names. For example, -c
makes the symbols ABC and abc equivalent.
. The -i option specifies a directory where the assembler can find included files such
as those following the .copy and .include directives.
2.3.3 Linker
The linker is used to combine multiple object files into a single executable program for
the target DSP hardware. It resolves external references and performs code relocation to
create the executable code. The C55x linker handles various requirements of different
object files and libraries as well as target system memory configurations. For a specific
hardware configuration, the system designers need to provide the memory mapping
specifications for the linker. This task can be accomplished by using a linker command
file. The Texas Instruments' visual linker is also a very useful tool that provides memory
usage directly.
The linker commands support expression assignment and evaluation, and provides
the MEMORY and SECTION directives. Using these directives, we can define the
memory configuration for the given target system. We can also combine object file
sections, allocate sections into specific memory areas, and define or redefine global
symbols at link time.
We can use the following command to invoke the C55x linker from the host system:
lnk55 [-options] filename_1,..., filename_n
The filename list (filename_1,..., filename_n) consists of object files created
by the assembler, linker command files, or achieve libraries. The default extension for
object files is .obj; any other extension must be explicitly specified. The options can
be placed anywhere on the command line to control different linking operations. For
example, the -o filename option can be used to specify the output executable file
name. If we do not provide the output file name, the default executable file name is
a.out. Some of the most common linker options are:
46 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
. The -ar option produces a re-locatable executable object file. The linker generates
an absolute executable code by default.
. The -e entry_point option defines the entry point for the executable module. This
will be the address of the first operation code in the program after power up or reset.
. The -stack size option sets the system stack size.
We can put the filenames and options inside the linker command file, and then invoke
the linker from the command line by specifying the command file name as follows:
lnk55 command_file.cmd
The linker command file is especially useful when we frequently invoke the linker with
the same information. Another important feature of the linker command file is that it
allows users to apply the MEMORY and SECTION directives to customize the pro-
gram for different hardware configurations. A linker command file is an ASCII text file
and may contain one or more of the following items:
. Input files (object files, libraries, etc.).
. Output files (map file and executable file).
. Linker options to control the linker as given from the command line of the shell
program.
. The MEMORY and SECTION directives define the target memory configuration
and information on how to map the code sections into different memory spaces.
The linker command file we used for the experiments in Chapter 1 is listed in Table
2.2. The first portion of the command file uses the MEMORY directive to identify the
range of memory blocks that physically exist in the target hardware. Each memory
block has a name, starting address, and block length. The address and length are given
in bytes. For example, the data memory is given a name called RAM, and it starts at the
byte address of hexadecimal 0x100, with a size of hexadecimal 0x1FEFF bytes.
The SECTIONS directive provides different code section names for the linker to
allocate the program and data into each memory block. For example, the program in
the .text section can be loaded into the memory block ROM. The attributes inside the
parenthesis are optional to set memory access restrictions. These attributes are:
R ± the memory space can be read.
W ± the memory space can be written.
X ± the memory space contains executable code.
I ± the memory space can be initialized.
There are several additional options that can be used to initialize the memory using
linker command files [2].
SOFTWARE DEVELOPMENT TOOLS 47
Table 2.2 Example of a linker command file used for the C55x simulator
/* Specify the system memory map */
MEMORY
{
RAM (RWIX) : origin ˆ 0100h, length ˆ 01FEFFh /* Data memory */
RAM2 (RWIX) : origin ˆ 040100h, length ˆ 040000h /* Program memory */
ROM (RIX) : origin ˆ 020100h, length ˆ 020000h /* Program memory */
VECS (RIX) : origin ˆ 0FFFF00h, length ˆ 00100h /* Reset vector */
}
/* Specify the sections allocation into memory */
SECTIONS
{
vectors > VECS /* Interrupt vector table */
.text > ROM /* Code */
.switch > RAM /* Switch table info */
.const > RAM /* Constant data */
.cinit > RAM2 /* Initialization tables */
.data > RAM /* Initialized data */
.bss > RAM /* Global & static variables */
.stack > RAM /* Primary system stack */
}
2.3.4 Code Composer Studio
As illustrated in Figure 2.8, the code composer studio (CCS) provides interface with the
C55x simulator (SIM), DSP starter kit (DSK), evaluation module (EVM), or in-circuit
emulator (XDS). The CCS supports both C and assembly programs.
The C55x simulator is available for PC and workstations, making it easy and
inexpensive to develop DSP software and to evaluate the performance of the processor
before designing any hardware. It accepts the COFF files and simulates the instructions
of the program such as the code running on the target DSP hardware. The C55x
simulator enables the users to single-step through the program, and observe the con-
tents of the CPU registers, data and I/O memory locations, and the current DSP states
of the status registers. The C55x simulator also provides profiling capabilities that tell
users the amount of time spent in one portion of the program relative to another. Since
all the functions of the TMS320C55x are performed on the host computer, the simula-
tion may be slow, especially for complicated DSP applications. Real world signals can
only be digitized and then later fed into a simulator as test data. In addition, the timing
of the algorithm under all possible input conditions cannot be tested using a simulator.
As introduced in Section 1.5, the various display windows and the commands of the
CCS provide most debugging needs. Through the CCS, we can load the executable object
code, display a disassembled version of the code along with the original source code, and
48 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
Code Composer Studio
.out .asm
.C
.lst/.map/.obj
lnk.cmd
Probe
file in
Probe
file out
Graphic
display
Profile
analysis
Program
debug
DSP
board
SIM
DSK
EVM
XDS
File
edit
Build
Siminit.cmd
Figure 2.8 TMS320C55x software development using CCS
view the contents of the registers and the memory locations. The data in the registers and
the memory locations can be modified manually. The data can be displayed in hexadeci-
mal, decimal integer, or floating-point formats. The execution of the program can be
single-stepped through the code, run-to-cursor, or controlled by applying breakpoints.
DSKand EVMare development boards with the C55x processor. They can be used for
real-time analysis of DSP algorithms, code logic verification, and simple application
tests. The XDS allows breakpoints to be set at a particular point in a program to examine
the registers and the memory locations in order to evaluate the real-time results using a
DSP board. Emulators allow the DSP software to run at full-speed in a real-time
environment.
2.3.5 Assembly Statement Syntax
The TMS320C55x assembly program statements may be separated into four ordered
fields. The basic syntax expression for a C55x assembly statement is
[label] [:] mnemonic [operand list] [;comment]
The elements inside the brackets are optional. Statements must begin with a label, blank,
asterisk, or semicolon. Each field must be separated by at least one blank. For ease of
reading and maintenance, it is strongly recommended that we use meaningful mnemonics
for labels, variables, and subroutine names, etc. An example of a C55x assembly state-
ment is shown in Figure 2.9. In this example, the auxiliary register, AR1, is initialized to a
constant value of 2.
Label field: A label can contain up to 32 alphanumeric characters (A±Z, a±z, 0±9, _ ,
and $). It associates a symbolic address with a unique program location. The line that is
labeled in the assembly program can then be referenced by the defined symbolic name.
This is useful for modular programming and branch instructions. Labels are optional,
but if used, they must begin in column 1. Labels are case sensitive and must start with an
alphabetic letter. In the example depicted in Figure 2.9, the symbol start is a label and
is placed in the first column.
Mnemonic field: The mnemonic field can contain a mnemonic instruction, an assem-
bler directive, macro directive, or macro call. The C55x instruction set supports both
SOFTWARE DEVELOPMENT TOOLS 49
my_symbol .set 2 ; my_symbol = 2
start mov #my_symbol,AR1 ; Load AR1 with 2
label
start at
column 1
mnemonic
mov
operand
src,dst
comments
begin with
a semicolon
Figure 2.9 An example of TMS320C55x assembly statement
DSP-specific operations and general-purpose applications (see the TMS320C55x DSP
Mnemonic Instruction Set Reference Guide [4] for details). Note that the mnemonic
field cannot start in column 1 otherwise it would be interpreted as a label. The
mnemonic instruction mov (used in Figure 2.9) copies the constant, my_symbol
(which is set to be 2 by .set directive) into the auxiliary register AR1.
Operand field: The operand field is a list of operands. An operand can be a constant, a
symbol, or a combination of constants and symbols in an expression. An operand can
also be an assembly-time expression that refers to memory, I/O ports, or pointers.
Another category of the operands can be the registers and accumulators. Constants
can be expressed in binary, decimal, or hexadecimal formats. For example, a binary
constant is a string of binary digits (0s and 1s) followed by the suffix B (or b) and a
hexadecimal constant is a string of hexadecimal digits (0, 1, . . . , 9, A, B, C, D, E, and F)
followed by the suffix H (or h). A hexadecimal number can also use a 0x prefix similar to
those used by C language. The prefix # is used to indicate an immediate constant. For
example, #123 indicates that the operand is a constant of decimal number 123, while
#0x53CD is the hexadecimal number of 53CD (equal to a decimal number of 21 453).
Symbols defined in an assembly program with assembler directives may be labels,
register names, constants, etc. For example, we use the .set directive to assign a
value to my_symbol in the example given by Figure 2.9. Thus the symbol my_symbol
becomes a constant of value during assembly time.
Comment field: Comments are notes about the program that are significant to the
programmer. A comment can begin with an asterisk or a semicolon in column one.
Comments that begin in any other column must begin with a semicolon.
2.4 TMS320C55x Addressing Modes
The TMS320C55x can address a total of 16 Mbytes of memory space. The C55x
supports the following addressing modes:
. Direct addressing mode
. Indirect addressing mode
50 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
. Absolute addressing mode
. Memory-mapped register addressing mode
. Register bits addressing mode
. Circular addressing mode
To explain the different addressing modes of the C55x, Table 2.3 lists the move
instruction (mov) with different syntax.
As illustrated in Table 2.3, each addressing mode uses one or more operands. Some of
the operand types are explained as follows:
. Smem means a data word (16-bit) from data memory, I/O memory, or MMRs.
. Lmem means a long data word (32-bit) from either data memory space or
MMRs.
. Xmem and Ymem are used by an instruction to perform two 16-bit data memory
accesses simultaneously.
. src and dst are source and destination registers, respectively.
. #k is a signed immediate 16-bit constant ranging from À32 768 to 32 767.
. dbl is a memory containing a long data word.
. xdst is an extended register (23-bit).
Table 2.3 C55x mov instruction with different operand forms
Instruction Description
1. mov #k, dst Load the 16-bit signed constant k to the destination
register dst
2. mov src, dst Load the content of source register src to the
destination register dst
3. mov Smem, dst Load the content of memory location Smem to the
destination register dst
4. mov Xmem, Ymem, ACx The content of Xmem is loaded into the lower part of
ACx while the content of Ymem is sign extended and
loaded into upper part of ACx
5. mov dbl(Lmem), pair(TAx) Load upper 16-bit data and lower 16-bit data from
Lmem to the TAx and TA(x‡1), respectively
6. amov #k23, xdst Load the effective address of k23 (23-bit constant) into
extended destination register (xdst)
TMS320C55X ADDRESSING MODES 51
2.4.1 Direct Addressing Mode
There are four types of direct addressing modes: data-page pointer (DP) direct, stack
pointer (SP) direct, register-bit direct, and peripheral data-page pointer (PDP) direct.
The DP direct mode uses the main data page specified by the 23-bit extended data-
page pointer (XDP). Figure 2.10 shows a generation of DP direct address. The upper
seven bits of the XDP (DPH) determine the main data page (0±127). The lower 16 bits
of the XDP (DP) define the starting address in the data page selected by the DPH. The
instruction contains the seven-bit offset in the data page (@x) that directly points to the
variable x (Smem). The data-page registers DPH, DP, and XDP can be loaded by the
mov instruction as
mov #k7, DPH ; Load DPH with a 7-bit constant k7
mov #k16, DP ; Load DP with a 16-bit constant k16
These instructions initialize the data pointer DPH and DP, respectively, using the
assembly code syntax, mov #k,dst, given in Table 2.3. The first instruction loads
the high portion of the extended data-page pointer, DPH, with a 7-bit constant k7 to set
up the main data page. The second instruction initializes the starting address of the
data-page pointer. The following is an example that initializes the DPH and DP
pointers:
Example 2.1: Instruction
mov #0x3, DPH
mov #0x0100, DP
DPH 0 DPH 03
DP 0000 DP 0100
Before instruction After instruction
The data-page pointer also can be initialized using a 23-bit constant as
amov #k23, XDP ; Load XDP with a 23-bit constant
This instruction initializes the XDP in one instruction. The syntax used in the assembly
code is given in Table 2.3, amov #k23, xdst, where #k23 is a 23-bit address and the
destination xdst is an extended register. The following example initializes the data-
page pointer XDP to data page 1 with starting address 0x4000:
DP (16 bits)
@x (7 bits)
DPH (7 bits)
DP direct address (23 bits)
+
XDP
Figure 2.10 The DP direct addressing mode to variable x
52 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
Example 2.2: Instruction
amov #0x14000, XDP
DPH 0 DPH 1
DP 0000 DP 4000
Before instruction After instruction
The following code details how to use DP direct addressing mode:
X .set 0x1FFEF
mov #0x1, DPH ; Load DPH with 1
mov #0x0FFEF, DP ; Load DP with starting address
.dp X
mov #0x5555, @X ; Store 0x5555 to memory location X
mov #0xFFFF, @(X+5) ; Store 0xFFFF to memory location X+5
In this example, the symbol @ tells the assembler that this access is using the direct
address mode. The directive .dp does not use memory space. It is used to indicate the
base address of the variable X.
The stack pointer (SP) direct addressing mode is similar to the DP direct addressing
mode. The 23-bit address can be formed with the extended stack pointer (XSP) in the
same way as the direct address that uses XDP. The upper seven bits (SPH) select the
main data page and the lower 16 bits (SP) determine the starting address of the stack
pointer. The 7-bit stack offset is contained in the instruction. When SPH ˆ 0 (main
page 0), the stack must not use the reserved memory space for MMRs from address 0 to
0x5F.
The I/O space address mode only has a 16-bit address range. The 512 peripheral data
pages are selected by the upper 9 bits of the PDP register. The 7-bit offset determines the
location inside the selected peripheral data page as illustrated in Figure 2.11.
2.4.2 Indirect Addressing Mode
Indirect addressing modes using index and displacement are the most powerful and
commonly used addressing modes. There are four types of indirect addressing
modes. The AR indirect mode uses one of the eight auxiliary registers as a pointer
to data memory, I/O space, and MMRs. The dual-AR indirect mode uses two
@x (7 bits)
Lower (7 bits) Upper (9 bits)
PDP direct address (16 bits)
+
PDP
Figure 2.11 The PDP direct addressing mode to variable x
TMS320C55X ADDRESSING MODES 53
auxiliary registers for dual data memory access. The coefficient data pointer
(CDP) indirect mode uses the CDP to point to data memory space. The coefficient-
dual-AR indirect mode uses the CDP and the dual-AR indirect modes for generating
three addresses. The coefficient-dual-AR indirect mode will be discussed later along
with pipeline parallelism.
The indirect addressing is the most frequently used addressing mode because it
provides powerful pointer update/modification schemes. Several pointer modification
schemes are listed in Table 2.4.
The AR indirect addressing mode uses an auxiliary register (AR0±AR7) to point to
data memory space. The upper seven-bit of the extended auxiliary register (XAR) points
to the main data page, while the lower 16-bit points to a data location on that page. Since
the I/O space address is limited to a 16-bit range, the upper portion of the XAR must be
set to zero when accessing I/O space. The next example uses indirect addressing mode,
where AR0 is used as the address pointer, and the instruction loads the data
content stored in data memory pointed by AR0 to the destination register AC0.
Example 2.3: Instruction
mov *AR0, AC0
AC0 00 0FAB 8678 AC0 00 0000 12AB
AR0 0100 AR0 0100
Data memory Data memory
0x100 12AB 0x100 12AB
Before instruction After instruction
Table 2.4 The AR and CDP indirect addressing pointer modifications
Operand ARn/CDP pointer modifications
*ARn or *CDP ARn (or CDP) is not modified.
*ARnÆ or
*CDPÆ
ARn (or CDP) is modified after the operation by:
Æ1 for 16-bit operation (ARnˆARnÆ1)
Æ2 for 32-bit operation (ARnˆARnÆ2)
*ARn(#k16)
or
*CDP(#k16)
ARn (or CDP) is not modified.
The signed 16-bit constant k16 is used as the offset for the base pointer
ARn (or CDP).
*‡ARn(#k16)
or
*‡CDP(#k16)
ARn (or CDP) is modified before the operation.
The signed 16-bit constant k16 is added as the offset to the base pointer
ARn (or CDP) before generating new address.
*(ARnÆT0/T1) ARn is modified after the operation by Æ16-bit content in T0 or T1,
(ARn ˆ ARnÆT0/T1)
*ARn(T0/T1) ARn is not modified.
T0 or T1 is used as the offset for the base pointer ARn.
54 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
The dual-AR indirect addressing mode allows two data-memory accesses through
the auxiliary registers AR0±AR7. It can access two 16-bit data in memory using the
syntax, mov Xmem, Ymem, ACx given in Table 2.3. The next example performs dual
16-bit data load with AR2 and AR3 as the data pointers to Xmem and Ymem, respect-
ively. The data pointed at by AR3 is sign-extended to 24-bit, loaded into the upper
portion of the destination register AC0(39:16), and the data pointed at by AR2
is loaded into the lower portion of AC0(15:0). The data pointers AR2 and AR3 are
also updated.
Example 2.4: Instruction
mov *AR2‡, *AR3À, AC0
AC0 FF FFAB 8678 AC0 00 3333 5555
AR2 0100 AR2 0101
AR3 0300 AR3 02FF
Data memory Data memory
0x100 5555 0x100 5555
0x300 3333 0x300 3333
Before instruction After instruction
The extended coefficient data pointer (XCDP) is the concatenation of the CDPH (the
upper 7-bit) and the CDP (the lower 16-bit). The CDP indirect addressing mode uses
the upper 7-bit to define the main data page and the lower 16-bit to point to the data
memory location within the specified data page. For the I/O space, only the 16-bit
address is used. An example of using the CDP indirect addressing mode is given as
follows:
Example 2.5: Instruction
mov *‡CDP(#2), AC3
AC3 00 0FAB EF45 AC3 00 0000 5631
CDP 0400 CDP 0402
Data memory Data memory
0x402 5631 0x402 5631
Before instruction After instruction
In this example, CDP is the pointer that contains the address of the coefficient in data
memory with an offset. This instruction increments the CDP pointer by 2 first, then
loads a coefficient pointed by the updated coefficient pointer to the destination register
AC3.
TMS320C55X ADDRESSING MODES 55
2.4.3 Absolute Addressing Mode
The memory can also be addressed using absolute addressing modes in either k16 or k23
absolute addressing modes. The k23 absolute mode specifies an address as a 23-bit
unsigned constant. The following example loads the data content at address 0x1234 on
main data page 1 into the temporary register, T2, where the symbol, *( ), represents the
absolute address mode.
Example 2.6: Instruction
mov *(#x011234), T2
T2 0000 T2 FFFF
Data memory Data memory
0x01 1234 FFFF 0x01 1234 FFFF
Before instruction After instruction
The k16 absolute addressing mode uses the operand *abs(#k16), where k16 is a
16-bit unsigned constant. The DPH (7-bit) is forced to 0 and concatenated with the
unsigned constant k16 to form a 23-bit data-space memory address. The I/O absolute
addressing mode uses the operand port(#k16). The absolute address can also be the
variable name such as the variable, x, in the following example:
mov *(x), AC0
This instruction loads the accumulator AC0 with a content of variable x. When using
absolute addressing mode, we do not need to worry about what is loaded into the data-
page pointer. The drawback of the absolute address is that it uses more code space to
represent the 23-bit address.
2.4.4 Memory-Mapped Register Addressing Mode
The absolute, direct, and indirect addressing modes introduced above can be used to
address MMRs. The MMRs are located in the data memory from address 0x0 to
0x5F on the main data page 0 as shown in Figure 2.6. To access the MMRs using
the k16 absolute operand, the DPH must be set to zero. The following example uses the
absolute addressing mode to load the 16-bit content of the AR2 into the temporary
register T2:
Example 2.7: Instruction
mov *abs16(#AR2), T2
AR2 1357 AR2 1357
T2 0000 T2 1357
Before instruction After instruction
56 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
For the MMR direct addressing mode, the DP addressing mode must be selected. The
example given next uses direct addressing mode to load the content of the lower portion of
the accumulator AC0(15:0), into the temporary register T0. When the mmap()qualifier
for the MMRdirect addressing mode is used, it forces the data address generator to act as
if the access is made to the main data page 0. That is, XDP ˆ 0.
Example 2.8: Instruction
mov mmap(@AC0L), T0
AC0 00 12DF 0202 AC0 00 12DF 0202
T0 0000 T0 0202
Before instruction After instruction
Accessing the MMRs using indirect addressing mode is the same as addressing the data
memory space. The address pointer can be either an auxiliary register or a CDP. Since
the MMRs are all located on data page 0, the XAR and XCDP must be initialized to
page 0 by setting all upper 7-bit to zero. The following instructions load the content of
AC0 into T1 and T2 temporary registers:
amov #AC0H, XAR6
mov *AR6À, T2
mov *AR6‡, T1
In this example, the first instruction loads the effective address of the upper portion of the
accumulator AC0 (AC0H, located at address 0x9 of page 0) to the extended auxiliary
register XAR6. That is, XAR6 ˆ 0x000009. The second instruction uses AR6 as a pointer
to copy the content of AC0H into the T2 register, and then the pointer decrements by 1 to
point to the lower portion of AC0 (AC0L, located at address 0x8 of page 0). The third
instruction copies the content of AC0L into the register T1 and modifies AR6 to point to
AC0H again.
2.4.5 Register Bits Addressing Mode
Both direct and indirect addressing modes can be used to address one bit or a pair of bits
of a specific register. The direct addressing mode uses a bit offset to access a particular
register's bit. The offset is the number of bits counting fromthe least significant bit (LSB),
i.e., bit 0. The bit test instruction will update the test condition bits, TC1 and TC2, of the
status register ST0. The instruction of register-bit direct addressing mode is shown in the
next example.
Example 2.9: Instruction
btstp @30, AC1
AC1 00 7ADF 3D05 AC1 00 7ADF 3D05
TC1 0 TC1 1
TC2 0 TC2 0
Before instruction After instruction
TMS320C55X ADDRESSING MODES 57
Using the indirect addressing modes to specify register bit(s) can be done as follows:
mov #2, AR4 ; AR4 contains the bit offset 2
bset *AR4, AC3 ; Set the AC3 bit pointed by AR4 to 1
btstp *AR4, AC1 ; Test AC1 bit-pair pointed by AR4
The register bit-addressing mode supports only the bit test, bit set, bit clear, and bit
complement instructions in conjunction with the accumulators (AC0±AC3), auxiliary
registers (AR0±AR7), and temporary registers (T0±T3).
2.4.6 Circular Addressing Mode
Circular addressing mode provides an efficient method for accessing data buffers
continuously without having to reset the data pointers. After accessing data, the data
buffer pointer is updated in a modulo fashion. That is, when the pointer reaches the
end of the buffer, it will wrap back to the beginning of the buffer for the next iteration.
Auxiliary registers (AR0±AR7) and the CDP can be used as circular pointers in
indirect addressing mode. The following steps are commonly used to set up circular
buffers:
1. Initialize the most significant 7-bit extended auxiliary register (ARnH or CDPH) to
select the main data page for a circular buffer. For example, mov #k7, AR2H.
2. Initialize the 16-bit circular pointer (ARn or CDP). The pointer can point to any
memory location within the buffer. For example, mov #k16, AR2 (the initialization
of the address pointer in the example of steps 1 and 2 can also be done using the
amov #k23, XAR2 instruction).
3. Initialize the 16-bit circular buffer starting address register (BSA01, BSA23, BSA45,
BSA67, or BSAC) associated with the auxiliary registers. For example, mov #k16,
BSA23, if AR2 (or AR3) is used as the circular addressing pointer register. The main
data page concatenated with the content of this register defines the 23-bit starting
address of the circular buffer.
4. Initialize the data buffer size register (BK03, BK47, or BKC). When using AR0±
AR3 (or AR4±AR7) as the circular pointer, BK03 (or BK47) should be initialized.
The instruction, mov #16, BK03, sets up a circular buffer of 16 elements for the
auxiliary registers, AR0±AR3.
5. Enable the circular buffer configuration by setting the appropriate bit in the status
register ST2. For example, the instruction bset AR2LC enables AR2 for circular
addressing.
Refer to the TMS320C55x DSP CPU Reference Guide [1] for details on circular
addressing mode. The following example demonstrates how to initialize a four integer
circular buffer, COEFF[4], and how the circular addressing mode accesses data in the
buffer:
58 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
amov #COEFF, XAR2 ; Main data page for COEFF[4]
mov #COEFF, BSA23 ; Buffer base address is COEFF[0]
mov #0x4, BK03 ; Set buffer size of 4-word
mov #2, AR2 ; AR2 points to COEFF[2]
bset AR2LC ; AR2 is configured as circular pointer
mov *AR2‡, T0 ; T0 is loaded with COEFF[2]
mov *AR2‡, T1 ; T1 is loaded with COEFF[3]
mov *AR2‡, T2 ; T2 is loaded with COEFF[0]
mov *AR2‡, T3 ; T3 is loaded with COEFF[1]
Since the circular addressing uses the indirect addressing modes, the circular pointers
can be updated using the modifications listed in Table 2.4. The use of circular buffers for
FIR filtering will be introduced in Chapter 5 in details.
2.5 Pipeline and Parallelism
The pipeline technique has been widely used by many DSP manufacturers to improve
processor performance. The pipeline execution breaks a sequence of operations into
smaller segments and executes these smaller pieces in parallel. The TMS320C55x uses
the pipelining mechanism to efficiently execute its instructions to reduce the overall
execution time.
2.5.1 TMS320C55x Pipeline
Separated by the instruction buffer unit, the pipeline operation is divided into two
independent pipelines ± the program fetch pipeline and the program execution pipeline
(see Figure 2.12). The program fetch pipeline consists of the following three stages (it
uses three clock cycles):
PA (program address): The C55x instruction unit places the program address on the
program-read address bus (PAB).
PM (program memory address stable): The C55x requires one clock cycle for its
program memory address bus to be stabilized before that memory can be read.
PB (program fetch from program data bus): In this stage, four bytes of the program
code are fetched from the program memory via the 32-bit program data-read bus (PB).
Figure 2.12 The C55x pipeline execution diagram
PIPELINE AND PARALLELISM 59
The code is placed into the instruction buffer queue (IBQ). For every clock cycle, the IU
will fetch four bytes to the IBQ. The numbers on the top of the diagram represent the
CPU clock cycle.
At the same time, the seven-stage execution pipeline performs the fetch, decode,
address, access, read, and execution sequence independent of the program fetch pipe-
line. The C55x program execution pipeline stages are summarized as follows:
F (fetch): In the fetch stage, an instruction is fetched from the IBQ. The size of the
instruction can be one byte for simple operations, or up to six bytes for more complex
operations.
D (decode): During the decoding process, decode logic gets one to six bytes from the
IBQ and decodes these bytes into an instruction or an instruction pair under the parallel
operation. The decode logic will dispatch the instruction to the program flow unit (PU),
address flow unit (AU), or data computation unit (DU).
AD (address): In this stage, the AU calculates data memory addresses using its data-
address generation unit (DAGEN), modifies pointers if required, and computes the
program-space address for PC-relative branching instructions.
AC (access cycles 1 and 2): The first cycle is used for the C55x CPU to send
the address for read operations to the data-read address buses (BAB, CAB, and
DAB), or transfer an operand to the CPU via the C-bus (CB). The second access
cycle is inserted to allow the address lines to be stabilized before the memory is read.
R (read): In the read stage, the data and operands are transferred to the CPU via the
CB for the Ymem operand, the B-bus (BB) for the Cmem operand, and the D-bus (DB)
for the Smem or the Xmem operands. For the Lmem operand read, both the CB and the
DB will be used. The AU will generate the address for the operand write and send the
address to the data-write address buses (EAB and FAB).
X (execute): Most data processing work is done in this stage. The ALU inside the AU
and the ALU inside the DU performs data processing execution, stores an operand via
the F-bus (FB), or stores a long operand via the E-bus and F-bus (EB and FB).
The C55x pipeline diagram illustrated in Figure 2.12 explains how the C55x pipeline
works. It is clear that the execution pipeline is full after seven cycles and every execution
cycle that follows will complete an instruction. If the pipeline is always full, this technique
increases the processing speed seven times. However, the pipeline flow efficiency is based
on the sequential execution of instruction. When a disturbing execution such as a branch
instruction occurs, the sudden change of the program flow breaks the pipeline sequence.
Under such circumstances, the pipeline will be flushed and will need to be refilled. This is
called pipeline break down. The use of IBQcan minimize the impact of the pipeline break
down. Proper use of conditional execution instructions to replace branch instructions can
also reduce the pipeline break down.
2.5.2 Parallel Execution
The parallelism of the TMS320C55x uses the processor's multiple-bus architecture, dual
MAC units, and separated PU, AU, and DU. The C55x supports two parallel process-
ing types ± implied and explicit. The implied parallel instructions are the built-in
instructions. They use the symbol of parallel columns, `: :', to separate the pair of
instructions that will be processed in parallel. The explicit parallel instructions are the
60 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
user-built instructions. They use the symbol of parallel bar, `j j', to indicate the pair of
parallel instructions. These two types of parallel instructions can be used together to
form a combined parallel instruction. The following examples show the user-built, built-
in, and combined parallel instructions. Each example is carried out in just one clock
cycle.
User-built:
mpym *AR1‡, *AR2‡, AC0 ; User-built parallel instruction
jjand AR4, T1 ; Using DU and AU
Built-in:
mac *AR0À, *CDPÀ, AC0 ; Built-in parallel instruction
::mac *AR1À, *CDPÀ, AC1 ; Using dual-MAC units
Built-in and User-built Combination:
mpy *AR2‡, *CDP‡, AC0 ; Combined parallel instruction
::mpy *AR3‡, *CDP‡, AC1 ; Using dual-MAC units and PU
j j rpt #15
Some of the restrictions when using parallel instructions are summarized as follows:
. For either the user-built or the built-in parallelism, only two instructions can be
executed in parallel, and these two instructions must not exceed six bytes.
. Not all instructions can be used for parallel operations.
. When addressing memory space, only the indirect addressing mode is allowed.
. Parallelism is allowed between and within execution units, but there cannot be any
hardware resources conflicts between units, buses, or within the unit itself.
There are several restrictions that define the parallelism within each unit when applying
parallelism to assembly coding. The detailed descriptions are given in the TMS320C55x
DSP Mnemonic Instruction Set Reference Guide [4].
The PU, AU, and DU can all be involved in parallel operations. Understanding the
register files in each of these units will help to be aware of the potential conflicts when
using the parallel instructions. Table 2.5 lists some of the registers in PU, AU, and
DU.
The parallel instructions used in the following example are incorrect because the
second instruction uses the direct addressing mode:
mov *AR2, AC0
j jmov T1, @x
We can correct the problem by replacing the direct addressing mode, @x, with an
indirect addressing mode, *AR1, so both memory accesses are using indirect addressing
mode as follows:
PIPELINE AND PARALLELISM 61
Table 2.5 Partial list of the C55x registers and buses
PU Registers/Buses AU Registers/Buses DU Registers/Buses
RPTC T0, T1, T2, T3 AC0, AC1, AC2, AC3
BRC0, BRC1 AR0, AR1, AR2, AR3, TRN0, TRN1
RSA0, RSA1 AR4, AR5, AR6, AR7
REA0, REA1 CDP
BSA01, BSA23, BSA45, BSA67
BK01, BK23, BK45, BK67
Read Buses: C, D Read Buses: C, D Read Buses: B, C, D
Write Buses: E, F Write Buses: E, F Write Buses: E, F
mov *AR2, AC0
j jmov T1, *AR1
Consider the following example where the first instruction loads the content of AC0
that resides inside the DU to the auxiliary register AR2 inside the AU. The second
instruction attempts to use the content of AC3 as the program address for a function
call. Because there is only one link between AU and DU, when both instructions try to
access the accumulators in the DU via the single link, it creates a conflict.
mov AC0, AR2
j jcall AC3
To solve the problem, we can change the subroutine call from call by accumulator to
call by address as follows:
mov AC0, AR2
j jcall my_func
This is because the instruction, call my_func, only needs the PU.
The coefficient-dual-AR indirect addressing mode is used to perform operations with
dual-AR indirect addressing mode. The coefficient indirect addressing mode supports
three simultaneous memory-accesses (Xmem, Ymem, and Cmem). The finite impulse
response (FIR) filter (will be introduced in Chapter 3) is an application that can
effectively use coefficient indirect addressing mode. The following code segment is an
example of using the coefficient indirect addressing mode:
mpy *AR1‡, *CDP‡, AC2 ; AR1 pointer to data X1
: :mpy *AR2‡, *CDP‡, AC3 ; AR2 pointer to data X2
| |rpt #6 ; Repeat the following 7 times
mac *AR1‡, *CDP‡, AC2 ; AC2 has accumulated result
: :mac *AR2‡, *CDP‡, AC3 ; AC3 has another result
In this example, the memory buffers (Xmem and Ymem) are pointed at by AR2 and AR3,
respectively, while the coefficient array is pointed at by CDP. The multiplication results
62 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
are added with the contents in the accumulators AC2 and AC3, and the final results are
stored back to AC2 and AC3.
2.6 TMS320C55x Instruction Set
We briefly introduced the TMS320C55x instructions and assembly syntax expression in
Section 2.3.5. In this section, we will introduce more useful instructions for DSP
applications. In general, we can divide the C55x instruction set into four categories:
arithmetic instructions, logic and bit manipulation instructions, load and store (move)
instructions, and program flow control instructions.
2.6.1 Arithmetic Instructions
Instructions used to performaddition (ADD), subtraction (SUB), and multiplication (MPY)
are arithmetic instructions. Most arithmetic operations can be executed conditionally.
The combination of these basic arithmetic operations produces another powerful subset
of instructions such as the multiply±accumulation (MAC) and multiply±subtraction
(MAS) instructions. The C55x also supports extended precision arithmetic such as
add-with-carry, subtract-with-borrow, signed/signed, signed/unsigned, and unsigned/
unsigned arithmetic instructions. In the following example, the multiplication instruc-
tion, mpym, multiplies the data pointed by AR1 and CDP, and the multiplication
product is stored in the accumulator AC0. After the multiplication, both pointers
(AR1 and CDP) are updated.
Example 2.10: Instruction
mpym *AR1‡, *CDPÀ, AC0
AC0 FF FFFF FF00 AC0 00 0000 0020
FRC 0 FRC 0
AR1 02E0 AR1 02E1
CDP 0400 CDP 03FF
Data memory Data memory
0x2E0 0002 0x2E0 0002
0x400 0010 0x400 0010
Before instruction After instruction
In the next example, the macmr40 instruction uses AR1 and AR2 as data pointers
and performs multiplication±accumulation. At the same time, the instruction also
carries out the following operations:
1. The key word `r' produces a rounded result in the high portion of the accumulator
AC3. After rounding, the lower portion of AC3(15:0) is cleared.
TMS320C55X INSTRUCTION SET 63
2. 40-bit overflow detection is enabled by the key word `40'. If overflow is detected,
the result in accumulator AC3 will be saturated to its 40-bit maximum value.
3. The option `T3 ˆ *AR1‡' loads the data pointed at by AR1 into the temporary
register T3 for later use.
4. Finally, AR1 and AR2 are incremented by one to point to the next data location in
memory space.
Example 2.11: Instruction
macmr40 T3ˆ *AR1‡, *AR2‡, AC3
AC3 00 0000 0020 AC3 00 235B 0000
FRC 1 FRC 1
T3 FFF0 T3 3456
AR1 0200 AR1 0201
AR2 0380 AR2 0381
Data memory Data memory
0x200 3456 0x200 3456
0x380 5678 0x380 5678
Before instruction After instruction
2.6.2 Logic and Bits Manipulation Instructions
Logic operation instructions such as AND, OR, NOT, and XOR (exclusive-OR) on data
values are widely used in program decision-making and execution flow control. They
are also found in many applications such as error correction coding in data commu-
nications. For example, the instruction and #0xf, AC0 clears all upper bits in the
accumulator AC0 but the four least significant bits.
Example 2.12: Instruction
and #0xf, AC0
AC0 00 1234 5678 AC0 00 0000 0008
Before instruction After instruction
The bit manipulation instructions act on an individual bit or a pair of bits of a register
or data memory. These types of instructions consist of bit clear, bit set, and bit test to a
specified bit (or a pair of bits). Similar to logic operations, the bit manipulation
instructions are often used with logic operations in supporting decision-making pro-
cesses. In the following example, the bit clear instruction clears the carry bit (bit 11) of
the status register ST0.
64 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
Example 2.13: Instruction
bclr #11, ST0
ST0 0800 ST0 0000
Before instruction After instruction
2.6.3 Move Instruction
The move instruction is used to copy data values between registers, memory locations,
register to memory, or memory to register. For example, to initialize the upper portion
of the 32-bit accumulator AC1 with a constant and zero out the lower portion of the
AC1, we can use the instruction mov #k(16, AC1, where the constant k is first shifted
left by 16-bit and then loaded into the upper portion of the accumulator AC1(31:16) and
the lower portion of the accumulator AC1(15:0) is zero filled. The 16-bit constant that
follows the # can be any signed number.
Example 2.14: Instruction
mov #5(16, AC1
AC1 00 0011 0800 AC1 00 0005 0000
Before instruction After instruction
A more complicated instruction completes the following several operations in one
clock cycle:
Example 2.15: Instruction
mov uns(rnd(HI(satuate(AC0(T2)))), *AR1‡
AC0 00 0FAB 8678 AC0 00 0FAB 8678
AR1 0x100 AR1 0x101
T2 0x2 T2 0x2
Data memory Data memory
0x100 1234 0x100 3EAE
Before instruction After instruction
1. The unsigned data content in AC0 is shifted to the left according to the content in
the temporary register T2.
2. The upper portion of the AC0(31:16) is rounded.
3. The data value in AC0 may be saturated if the left-shift or the rounding process
causes the result in AC0 to overflow.
TMS320C55X INSTRUCTION SET 65
4. The final result after left shifting, rounding, and maybe saturation, is stored into the
data memory pointed at by the pointer AR1.
5. Pointer AR1 is automatically incremented by 1.
2.6.4 Program Flow Control Instructions
The program flow control instructions are used to control the execution flow of the
program, including branching (B), subroutine call (CALL), loop operation (RPTB),
return to caller (RET), etc. All these instructions can be either conditionally or uncondi-
tionally executed. For example,
callcc my_routine, TC1
is the conditional instruction that will call the subroutine my_routine only if the test
control bit TC1 of the status register ST0 is set. Conditional branch (BCC) and condi-
tional return (RETCC) can be used to control the program flow according to certain
conditions.
The conditional execution instruction, xcc, can be implemented in either condi-
tional execution or partial conditional execution. In the following example, the
conditional execution instruction tests the TC1 bit. If TC1 is set, the instruction,
mov *AR1‡, AC0, will be executed, and both AC0 and AR1 are updated. If the
condition is false, AC0 and AR1 will not be changed. Conditional execution instruction
xcc allows for the conditional execution of one instruction or two paralleled instruc-
tions. The label is used for readability, especially when two parallel instructions are
used.
Example 2.16: Instruction
xcc label, TC1
mov *AR1‡, AC0
label
TC1 ˆ 1 TC1 ˆ 0
AC0 00 0000 0000 AC0 00 0000 55AA AC0 00 0000 0000 AC0 00 0000 0000
AR1 0x100 AR1 0x101 AR1 0x100 AR1 0x100
Data memory Data memory Data memory Data memory
0x100 55AA 0x100 55AA 0x100 55AA 0x100 55AA
Before instruction After instruction Before instruction After instruction
In addition to conditional execution, the C55x also provides the capability of partially
conditional execution of an instruction. An example of partial conditional execution is
given as follows:
66 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
Example 2.17: Instruction
xccpart label, TC1
mov *AR1‡, AC0
label
TC1 ˆ 1 TC1 ˆ 0
AC0 00 0000 0000 AC0 00 0000 55AA AC0 00 0000 0000 AC0 00 0000 0000
AR1 0x100 AR1 0x101 AR1 0x100 AR1 0x101
Data memory Data memory Data memory Data memory
0x100 55AA 0x100 55AA 0x100 55AA 0x100 55AA
Before instruction After instruction Before instruction After instruction
When the condition is true, both AR1 and AC0 will be updated. However, if the
condition is false, the execution phase of the pipeline will not be carried out. Since the
first operand (the address pointer AR1) is updated in the read phase of the pipeline, AR1
will be updated whether or not the condition is true, while the accumulator AC0 will
remain unchanged at the execution phase. That is, the instruction is only partially
executed.
Many real-time DSP applications require repeated executions of some instructions
such as filtering processes. These arithmetic operations may be located inside nested
loops. If the number of data processing instructions in the inner loop is small, the
percentage of overhead for loop control may be very high. The loop control instruc-
tions, such as testing and updating the loop counter(s), pointer(s), and branches back to
the beginning of the loop to execute the loop again, impose a heavy overhead for the
processor. To minimize the loop overhead, the C55x includes built-in hardware for zero-
overhead loop operations.
The single-repeat instruction(RPT) repeats the following single-cycle instruction or two
single-cycle instructions that are executed in parallel. For example,
rpt #NÀ1 ; Repeat next instruction N times
instruction_A
The number, NÀ1, is loaded into the single-repeat counter (RPTC) by the RPT
instruction. The following instruction_A will be executed N times.
The block-repeat instruction (RPTB) forms a loop that repeats a block of instructions.
It supports a nested loop with an inner loop being placed inside an outer loop. Block-
repeat registers use block-repeat counters BRC0 and BRC1. For example,
mov #NÀ1, BRC0 ; Repeat outer loop N times
mov #MÀ1, BRC1 ; Repeat inner loop M times
rptb outloop-1 ; Repeat outer loop up to outloop
mpy *AR1‡, *CDP‡, AC0
mpy *AR2‡, *CDP‡, AC1
rptb inloop-1 ; Repeat inner loop up to inloop
mac *AR1‡, *CDP‡, AC0
mac *AR2‡, *CDP‡, AC1
inloop ; End of inner loop
TMS320C55X INSTRUCTION SET 67
mov AC0, *AR3‡ ; Save result in AC0
mov AC1, *AR4‡ ; Save result in AC1
outloop ; End of outer loop
The above example uses two repeat instructions to control a nested repetitive oper-
ation. The block-repeat structure
rptb label_name-1
(more instructions...)
label_name
executes a block of instructions between the rptb instruction and the end label
label_name. The maximum number of instructions that can be used inside a block-
repeat loop is limited to 64 Kbytes of code. Because of the pipeline scheme, the minimum
cycles within a block-repeat loop are two. The maximum number of times that a loop can
be repeated is limited to 65 536 (ˆ 2
16
) because of the 16-bit block-repeat counters.
2.7 Mixed C and Assembly Language Programming
As discussed in Chapter 1, the mixing of C and assembly programs are used for many
DSP applications. C code provides the ease of maintenance and portability, while
assembly code has the advantages of run-time efficiency and code density. We can
develop C functions and assembly routines, and use them together. In this section, we
will introduce how to interface C with assembly programs and review the guidelines of
the C function calling conventions for the TMS320C55x.
The assembly routines called by a C function can have arguments and return values
just like C functions. The following guidelines are used for writing the C55x assembly
code that is callable by C functions.
Naming convention: Use the underscore `_' as a prefix for all variables and routine
names that will be accessedby Cfunctions. For example, use _my_asm_funcas the name
of an assembly routine called by a C function. If a variable is defined in an assembly
routine, it must use the underscore prefix for Cfunctiontoaccess it, suchas _my_var. The
prefix `_' is used by the Ccompiler only. When we access assembly routines or variables in
C, we don't need to use the underscore as a prefix. For example, the following C program
calls the assembly routine using the name my_asm_func without the underscore:
extern int my_asm_func /* Reference an assembly function */
void main()
{
int c; /* Define local variable */
c ˆ my_asm_func(); /* Call the assembly function */
}
This C program calls the following assembly routine:
.global _my_asm_func ; Define the assembly function
_my_asm_func ; Name of assembly routine
mov #0x1234, T0
ret ; Return to the call function
68 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
Variable definition: The variables that are accessed by both C and assembly routines
must be defined as global variables using the directive .global, .def, or .ref by the
assembler.
Compiler mode: By using the C compiler, the C55x CPL (compiler mode) bit is
automatically set for using stack-pointer (SP) relative addressing mode when entering
an assembly routine. The indirect addressing modes are preferred under this configur-
ation. If we need to use direct addressing modes to access data memory in a C callable
assembly routine, we must change to DP (data-page) direct addressing mode. This can
be done by clearing the CPL bit. However, before the assembly routine returns to its C
caller function, the CPL bit must be restored. The bit clear and bit set instructions,
bclr CPL and bset CPL, can be used to reset and set the CPL bit in the status register
ST1, respectively. The following code can be used to check the CPL bit, turn CPL bit off
if it is set, and restore the CPL bit before returning it to the caller.
btstclr #14, *(ST1), TC1 ; Turn off CPL bits if it is set
(more instructions...)
xcc continue, TC1 ; TC1 is set if we turned CPL bit off
bset CPL ; Turn CPL bit on
continue
ret
Passing arguments: To pass arguments from a C function to an assembly routine, we
must follow the strict rules of C-callable conversions set by the C compiler. When
passing an argument, the C compiler assigns it to a particular data type and then places
it in a register according to its data type. The C55x C compiler uses the following three
classes to define the data types:
. Data pointer: int *, or long *.
. 16-bit data: char, short, or int.
. 32-bit data: long, float, double, or function pointers.
If the arguments are pointers to data memory, they are treated as data pointers. If the
argument can fit into a 16-bit register such as int and char, it is considered to be 16-bit
data. Otherwise, it is considered 32-bit data. The arguments can also be structures. A
structure of two words (32 bits) or less is treated as a 32-bit data argument and is passed
using a 32-bit register. For structures larger than two words, the arguments are passed
by reference. The C compiler will pass the address of a structure as a pointer, and this
pointer is treated like a data argument.
For a subroutine call, the arguments are assigned to registers in the order that the
arguments are listed in the function. They are placed in the following registers according
to their data type, in an order shown in Table 2.6.
Note in Table 2.6 the overlap between AR registers used for data pointers and the
registers used for 16-bit data. For example, if T0 and T1 hold 16-bit data arguments,
and AR0 already holds a data pointer argument, a third 16-bit data argument would be
placed into AR1. See the second example in Figure 2.13. If the registers of the appro-
priate type are not available, the arguments are passed onto the stack. See the third
example in Figure 2.13.
MIXED C AND ASSEMBLY LANGUAGE PROGRAMMING 69
Table 2.6 Argument classes assigned to registers
Argument type Register assignment order
16-bit data pointer AR0, AR1, AR2, AR3, AR4
23-bit data pointer XAR0, XAR1, XAR2, XAR3, XAR4
16-bit data T0, T1, AR0, AR1, AR2, AR3, AR4
32-bit data AC0, AC1, AC2
T0 T0 AC0 AR0
↓ ↓ ↓ ↓
int func(int i1, long l2, int *p3);
AC0 AR0 T0 T1 AR1
↓ ↓ ↓ ↓ ↓
long func (int *p1, int i2, int i3, int i4);
AC0 AC1 AC2 Stack T0
↓ ↓ ↓ ↓ ↓
void func (long l1, long l2, long l3, long l4, int i5);
Figure 2.13 Examples of arguments passing conventions
Return values: The calling function/routine collects the return value from the called
function/subroutine. A 16-bit integer data is returned by the register T0 and a 32-bit
data is returned in the accumulator AC0. A data pointer is returned in (X)AR0. When
the called routine returns a structure, the structure is on the local stack.
Register use and preservation: When making a function call, the register assignments
and preservations between the caller and called functions are strictly defined. Table 2.7
describes how the registers are preserved during a function call. The called function
must save the contents of the save-on-entry registers (T2, T3, AR5, AR6, and AR7) if
it will use these registers. The calling function must push the contents of any other
save-on-call registers onto the stack if these register's contents are needed after the
function/subroutine call. A called function can freely use any of the save-on-call
registers (AC0±AC2, T0, T1, and AR0±AR4) without saving its value. More detailed
descriptions can be found in the TMS320C55x Optimizing C Compiler User's
Guide [3].
2.8 Experiments ± Assembly Programming Basics
We have introduced the TMS320C55x assembly language and several addressing modes.
Experiments given in this section will help to use different addressing modes for writing
assembly code. We also introduced the C function interfacing with assembly routines
and we will explore the C-assembly interface first.
70 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
Table 2.7 Register use and preservation conventions
Registers Preserved by Used for
AC0±AC2 Calling function
Save-on-call
16, 32, or 40-bit data
24-bit code pointers
(X)AR0±(X)AR4 Calling function
Save-on-call
16-bit data
16 or 23-bit pointers
T0 and T1 Calling function
Save-on-call
16-bit data
AC3 Called function
Save-on-entry
16, 32, or 40-bit data
(X)AR5±(X)AR7 Called function
Save-on-entry
16-bit data
16 or 23-bit pointers
T2 and T3 Called function
Save-on-entry
16-bit data
2.8.1 Experiment 2A ± Interfacing C with Assembly Code
In this experiment, we will learn how to write C-callable assembly routines. The
following example illustrates a C function main, which calls an assembly routine to
perform a summation, and returns the result back to the C main function. The C
program exp2a.c is listed as follows:
extern int sum(int *); /* Assembly routine sum */
int x[2] ˆ {0x1234, 0x4321}; /* Define x[] as global array */
int s; /* Define s as global variable */
void main()
{
s ˆ sum(x); /* Call assembly routine _sum */
}
The assembly routine exp2_sum.asm is listed as follows:
.global _sum
_sum
mov *AR0‡, AC0 ; AC0 ˆ x[1]
add *AR0‡, AC0 ; AC0 ˆ x[1]‡x[2]
mov AC0, T0 ; Return value in T0
ret ; Return to calling function
where the label _sum defines the starting or entry of the assembly subroutine and
directive .global defines that the assembly routine _sum as a global function.
Perform the following steps for Experiment 2A:
1. Use the CCS to create a project called exp2a in A:\Experiment2.
EXPERIMENTS ± ASSEMBLY PROGRAMMING BASICS 71
2. Write exp2a.c based on the C sample code given above, save it in
A: \Experiment2.
3. Write exp2_sum.asm based on the assembly sample code given above and save it
in A: \Experiment2.
4. Copy the linker command file, exp1.cmd, from previous experiment, rename it to
exp2.cmd and save it to A: \Experiment2.
5. From the CCS Project-Options-Linker-Library tab, to include the run-time
4. support library rst55.lib.
6. From CCS Project-Options-Compiler, sets include file search path if necessary.
4. Under the Category-Assembly, check the Keep .asm files box and interlist C and
ASM statements box.
7. Compile and debug the code, then load exp2a.out and issue the Go-Main
command.
8. Watch and record the changes in the AC0, AR0, and T0 in the CPU register
window. Watch memory location `s' and `x' and record when the content of result
`s' is updated. Why?
9. Single-step through the C and assembly code.
10. Examine the assembly code exp2a.asm that the C compiler generated. How is the
return value passed to the C calling function?
11. If we define the result `s' inside the function main(), from which location can we
view its value? Why?
2.8.2 Experiment 2B ± Addressing Mode Experiments
In Section 2.4, we introduced six C55x addressing modes. In the second part of the
experiments, we will write assembly routines to exercise different addressing modes to
understand how each of these addressing modes work. The assembly routines for this
experiment are called by a C function.
1. Write a C function called exp2b.c as follows:
int Ai[8]; /* Define array Ai[] */
int Xi[8]; /* Define array Xi[] */
int result1, result2; /* Define variables */
main( )
{
void exp2b_1(void);
void exp2b_2(void);
72 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
result1 ˆ exp2b_3(Ai, Xi);
result2 ˆ exp2b_4(Ai, Xi);
}
2. Save the file as exp2b.c in A: \Experiment2.
2. Write assembly routine void exp2b_1(void)using the absolute addressing mode
to initialize the array Ai[8]ˆ {1, 2, 3, 4, 5, 6, 7, 8}in data memory as in the
following example:
.global _Ai
mov #1, *(_Ai)
mov #2, *(_Ai‡1)
(more instructions...)
2. Since the array Ai[8]is defined in the C function, the assembly routine references it
using the directive .global (or .ref).
3. Write the assembly routine void exp2b_2(void)using the direct addressing
mode to initialize the array X[8]ˆ {9, 3, 2, 0, 1, 9, 7, 1}in data memory.
As we mentioned in Section 2.4.1, there are four different direct addressing modes.
We use the DP direct addressing mode for this part of the experiment. The data-
page pointer, XDP, needs to be set before we can start using the direct addressing
mode. The Xi array can be initialized using the following assembly code:
btstclr #14, *(ST1), TC1 ; Clear CPL bit for DP addressing
amov #_Xi, XDP ; Initialize XDP to point to Xi
.dp _Xi
mov #9, @_Xi ; Using direct addressing mode
mov #3, @_Xi‡1
(more instructions...)
xcc continue, TC1
bset CPL ; Reset CPL bit if it was cleared
continue
2. The instruction btstclr #14, *(ST1), TC1 tests the CPL (compiler mode) bit (bit
14) of the status register ST1. The compiler mode will be set if this routine is called
by a C function. If the test is true, the test flag bit, TC1 (bit 13 of status register ST0)
will be set, and the instruction clears the CPL bit. This is necessary for using DP
direct addressing mode instead of SP direct addressing mode. At the end of the code
section, the conditional execution instruction, xcc continue, TC1, is used to set
the CPL bit if TC1 was set.
4. The sum-of-product operation (dot product) is one of the most commonly used
functions by DSP systems. A dot product of two vectors of length L can be
expressed as
Y ˆ
X
LÀ1
iˆ0
A
i
X
i
ˆ A
0
X
0
‡A
1
X
1
‡Á Á Á ‡A
LÀ1
X
LÀ1
;
EXPERIMENTS ± ASSEMBLY PROGRAMMING BASICS 73
2. where the vectors A
i
and X
i
are one-dimensional arrays of length L. There are many
different ways to access the elements of the arrays, such as direct, indirect, and
absolute addressing modes. In the following experiment, we will write a subroutine
int exp2b_3(int *Ai, int *Xi)to perform the dot product using indirect
addressing mode, and store the returned value in the variable result in data
memory. The code example is given as follows:
; Assume AR0 and AR1 are pointing to Ai and Xi
mpym *AR0‡, *AR1‡, AC0 ; Multiply Ai[0]and Xi[0]
mpym *AR0‡, *AR1‡, AC1 ; Multiply Ai[1]and Xi[1]
add AC1, AC0 ; Accumulate the partial result
mpym *AR0‡, *AR1‡, AC1 ; Multiply Ai[2]and Xi[2]
add AC1, AC0 ; Accumulate the partial result
(more instructions...)
mov AC0, T0
4. In the program, arrays A
i
and X
i
are defined as global arrays in the exp2.c. The
A
i
and X
i
arrays have the same data values as given previously. The return value is
passed to the calling function by T0.
5. Write an assembly routine int exp2b_4(int *Ai, int *Xi)using the indirect
addressing mode in conjunction with parallel instructions and repeat instructions
to improve the code density and efficiency. The following is an example of the
code:
mpym *AR0‡, *AR1‡, AC0 ; Multiply Ai[0]and Xi[0]
j jrpt #6 ; Multiply and accumulate the rest
macm *AR0‡, *AR1‡, AC0
4. The auxiliary registers, AR0 and AR1, are used as data pointers to array A
i
and
array X
i
, respectively. The instruction macm performances multiply-and-accumu-
late operation. The parallel bar jj indicates the parallel operation of two instruc-
tions. The repeat instruction, rpt #K will repeat the following instruction K‡1
times.
6. Create a project called exp2b and save it in A: \Experiment2.
7. Use exp2.cmd, exp2b.c, exp2b_1.asm, exp2b_2.asm, exp2b_3 .asm, and
exp2b_4.asm to build the project.
8. Open the memory watch window to watch how the arrays A
i
and X
i
are initialized
in data memory by the assembly routine exp2b_1.asm and exp2b_2.asm.
9. Open the CPU registers window to see how the dot product is computed by
exp2b_3.asm, and exp2b_4.asm.
10. Use the profile capability learned from the experiments given in Chapter 1 to
measure the run-time of the sum-of-product operations and compare the cycle
difference of the routine exp2b_3.asm and exp2b_4.asm.
74 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
11. Use the map file to compare the assembly program code size of routine
exp2b_3.asm and exp2b_4.asm. Note that the program size is given in bytes.
References
[1] Texas Instruments, Inc., TMS320C55x DSP CPU Reference Guide, Literature no. SPRU371A,
2000.
[2] Texas Instruments, Inc., TMS320C55x Assembly Language Tools User's Guide, Literature no.
SPRU380, 2000.
[3] Texas Instruments, Inc., TMS320C55x Optimizing C Compiler User's Guide, Literature no.
SPRU281, 2000.
[4] Texas Instruments, Inc., TMS320C55x DSP Mnemonic Instruction Set Reference Guide, Literature
no. SPRU374A, 2000.
[5] Texas Instruments, Inc., TMS320C55x DSP Algebraic Instruction Set Reference Guide, Literature
no. SPRU375, 2000.
[6] Texas Instruments, Inc., TMS320C55x Programmer's Reference Guide, Literature no. SPRU376,
2000.
Exercises
1. Check the following examples to determine if these are correct parallel instructions. If not,
correct the problems.
(a) mov *AR1‡, AC1
: : add @x, AR2
(b) mov AC0, dbl(*AR2‡)
: : mov dbl(*(AR1‡T0)), AC2
(c) mpy *AR1‡, *AR2‡, AC0
: : mpy *AR3‡, *AR2‡, AC1
j j rpt #127
2. Given a memory block and XAR0, XDP, and T0 as shown in Figure 2.14. Determine the
contents of AC0, AR0, and T0 after the execution of the following instructions:
(a) mov *(#x‡2), AC0
(b) mov @(xÀx‡1), AC0
(c) mov @(xÀx‡0x80), AC0
(d) mov *AR0‡, AC0
(e) mov *(AR0‡T0), AC0
(f) mov *AR0(T0), AC0
(g) mov *AR0(#À1), AC0
(h) mov *AR0(#2), AC0
(i) mov *AR0(#0x80), AC0
3. C functions are defined as follows. Use Table 2.8 to show how the C compiler passes
parameters for each of the following functions:
EXERCISES 75
0xFFFF
0x0000
0x1111
0x2222
0x3333
0x4444
0x00FFFF
x = 0x010000
0x010001
0x010002
0x010003
0x010004
data
memory
address:
0x8080 0x010080
:
:
:
:
0x0004 T0
0x010000 XDP
0x010000 XAR0
Figure 2.14 Data memory and registers
(a) int func_a(long, int, int, int, int, int,
int *, int *, int *, int *);
var ˆ func_a(0xD32E0E1D, 0, 1, 2, 3, 4, pa, pb, pc, pd);
(b) int func_b(long, long, long, int, int, int,
int *, int *, int *, int *);
var ˆ func_b(0x12344321, 0, 1, 2, 3, 4, pa, pb, pc, pd);
(c) long func_c(int, int *, int *, int, int, long,
int *, int *, long, long);
var ˆ func_c(0x2468, pa, pb, 1, 2, 0x1001,
pc, pd, 0x98765432, 0x0);
Table 2.8 List of parameters passed by the C55x C compiler
T0 T1 T2 T3 AC0 AC1 AC2 AC3
XAR0 XAR1 XAR2 XAR3 XAR4 XAR5 XAR6 XAR7
SP(À3) SP(À2) SP(À1) SP(0) SP(1) SP(2) SP(3) var
76 INTRODUCTION TO TMS320C55X DIGITAL SIGNAL PROCESSOR
3
DSP Fundamentals
and Implementation
Considerations
The derivation of discrete-time systems is based on the assumption that the signal and
system parameters have infinite precision. However, most digital systems, filters, and
algorithms are implemented on digital hardware with finite wordlength. Therefore DSP
implementation with fixed-point hardware requires special attention because of the
potential quantization and arithmetic errors, as well as the possibility of overflow.
These effects must always be taken into consideration in DSP system design and
implementation.
This chapter presents some fundamental DSP concepts in time domain and practical
considerations for the implementation of digital filters and algorithms on DSP hard-
ware. Sections 3.1 and 3.2 briefly review basic time-domain DSP issues. Section 3.3
introduces probability and random processes, which are useful in analyzing the finite-
precision effects in the latter half of the chapter and adaptive filtering in Chapter 8. The
rigorous treatment of these subjects can be found in other DSP books listed in the
reference. Readers who are familiar with these DSP fundamentals should be able to skip
through some of these sections. However, most notations used throughout the book will
be defined in this chapter.
3.1 Digital Signals and Systems
In this section, we will define some widely used digital signals and simple DSP systems.
The purpose of this section is to provide the necessary background for understanding
the materials presented in subsequent sections and later chapters.
3.1.1 Elementary Digital Signals
There are several ways to describe signals. For example, signals encountered in com-
munications are classified as deterministic or random. Deterministic signals are used
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
for testing purposes and for mathematically describing certain phenomena. Random
signals are information-bearing signals such as speech. Some deterministic signals will be
introduced in this section, while random signals will be discussed in Section 3.3.
As discussed in Chapter 1, a digital signal is a sequence of numbers
¦x(n), ÷· < n < ·¦, where n is the time index. The unit-impulse sequence, with
only one non-zero value at n = 0, is defined as
d(n) =
1, n = 0
0, n = 0,

(3:1:1)
where d(n) is also called the Kronecker delta function. This unit-impulse sequence is
very useful for testing and analyzing the characteristics of DSP systems, which will be
discussed in Section 3.1.3.
The unit-step sequence is defined as
u(n) =
1, n _ 0
0, n < 0:

(3:1:2)
This signal is very convenient for describing a causal (or right-sided) signal x(n)
for n _ 0. Causal signals are the most commonly encountered signals in real-time
DSP systems.
Sinusoidal signals (sinusoids or sinewaves) are the most important sine (or cosine)
signals that can be expressed in a simple mathematical formula. They are also good
models for real-world signals. The analog sinewave can be expressed as
x(t) = Asin(Vt ÷f) = Asin(2pft ÷f), (3:1:3)
where A is the amplitude of the sinewave,
V = 2pf (3:1:4)
is the frequency in radians per second (rad/s), f is the frequency in Hz, and f is the
phase-shift (initial phase at origin t = 0) in radians.
When the analog sinewave defined in (3.1.3) is connected to the DSP system shown in
Figure 1.1, the digital signal x(n) available for the DSP hardware is the causal sinusoidal
signal
x(n) = Asin(VnT ÷f), n = 0, 1, . . . , ·
= Asin(VnT ÷f)u(n)
= Asin(2pfnT ÷f)u(n),
(3:1:5)
where T is the sampling period in seconds. This causal sequence can also be expressed
as
x(n) = Asin(!n ÷f)u(n) = Asin(2pFn ÷f)u(n) , (3:1:6)
78 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
where
! = VT =
V
f
s
(3:1:7)
is the discrete-time frequency in radians per sample and
F = f T =
f
f
s
(3:1:8)
is the normalized frequency to its sampling frequency, f
s
, in cycles per sample.
The fundamental difference between describing the frequency of analog and digital
signals is summarized in Table 3.1. Analog signal sampling implies a mapping of an
infinite range of real-world frequency variable f (or V) into a finite range of discrete-
time frequency variable F (or !). The highest frequency in a digital signal is F = 1=2 (or
! = p) based on Shannon's sampling theorem defined in (1.2.3). Therefore the spectrum
of discrete-time (digital) signals is restricted to a limited range as shown in Table 3.1.
Note that some DSP books define the normalized frequency as F = f

( f
s
=2) with
frequency range ÷1 _ F _ 1.
Example 3.1: Generate 64 samples of a sine signal with A = 2, f = 1000 Hz, and
f
s
= 8 kHz using MATLAB. Since F = f =f
s
= 0:125, we have ! = 2pF = 0:25p.
From Equation (3.1.6), we need to generate x(n) = 2 sin(!n), for n = 0, 1, . . . , 63.
These sinewave samples can be generated and plotted by the following MATLAB
script:
n = [0:63];
omega = 0.25*pi;
xn = 2*sin(omega*n);
plot(n, xn);
3.1.2 Block Diagram Representation of Digital Systems
A DSP system (or algorithm) performs prescribed operations on digital signals. In some
applications, we view a DSP system as an operation performed on an input signal, x(n),
in order to produce an output signal, y(n), and express the general relationship between
x(n) and y(n) as
Table 3.1 Units, relationships, and range of four frequency variables
Variables Unit Relationship Range
V radians per second V = 2pf ÷· < V < ·
f cycles per second (Hz) f =
F
T
÷· < f < ·
! radians per sample ! = 2pF ÷p _ ! _ p
F cycles per sample F =
f
f
s
÷
1
2
_ F _
1
2
DIGITAL SIGNALS AND SYSTEMS 79
y(n) = T[x(n)[, (3:1:9)
where T denotes the computational process for transforming the input signal, x(n), into
the output signal, y(n). A block diagram of the DSP system defined in (3.1.9) is
illustrated in Figure 3.1.
The processing of digital signals can be described in terms of combinations of certain
fundamental operations on signals. These operations include addition (or subtraction),
multiplication, and time shift (or delay). A DSP system consists of the interconnection
of three basic elements ± adders, multipliers, and delay units.
Two signals, x
1
(n) and x
2
(n), can be added as illustrated in Figure 3.2, where
y(n) = x
1
(n) ÷x
2
(n) (3:1:10)
is the adder output. With more than two inputs, the adder could be drawn as a multi-
input adder, but the additions are typically done two inputs at a time in digital hard-
ware. The addition operation of Equation (3.1.10) can be implemented as the following
C55x code using direct addressing mode:
mov @x1n, AC0 ; AC0 = x1(n)
add @x2n, AC0 ; AC0 = x1(n)÷x2(n) = y(n)
A given signal can be multiplied by a constant, a, as illustrated in Figure 3.3, where
x(n) is the multiplier input, a represents the multiplier coefficient, and
y(n) = ax(n) (3:1:11)
x(n)
DSP system
T[ ]
y(n) = T[x(n)]
Figure 3.1 Block diagram of a DSP system
x
1
(n) x
1
(n)
x
2
(n) y(n) x
2
(n) y(n)
+
+
Σ
or
Figure 3.2 Block diagram of an adder
x(n) y(n)
a
or
x(n) y(n) a
Figure 3.3 Block diagram of a multiplier
80 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
x(n) y(n) = x(n−1)
z
−1
Figure 3.4 Block diagram of a unit delay
is the multiplier's output. The multiply operation of equation (3.1.11) can be imple-
mented by the following C55x code using indirect addressing mode:
amov #alpha, XAR1 ; AR1 points to alpha ()
amov #xn, XAR2 ; AR2 points to x(n)
mpy *AR1, *AR2, AC0 ; AC0 = *x(n) = y(n)
The sequence {x(n)} can be shifted (delayed) in time by one sampling period, T, as
illustrated in Figure 3.4. The box labeled z
÷1
represents the unit delay, x(n) is the input
signal, and
y(n) = x(n ÷1) (3:1:12)
is the output signal, which is the input signal delayed by one unit (a sampling period).
In fact, the signal x(n ÷1) is actually the stored signal x(n) one sampling period
(T seconds) before the current time. Therefore the delay unit is very easy to implement
in a digital system, but is difficult to implement in an analog system. A delay by more
than one unit can be implemented by cascading several delay units in a row. Therefore
an L-unit delay requires L memory locations configured as a first-in first-out buffer,
which can also be implemented as a circular buffer (will be discussed in Chapter 5) in
memory.
There are several ways to implement delay operations on the TMS320C55x. The
following code uses a delay instruction to move the contents of the addressed data
memory location into the next higher address location:
amov #xn, XAR1 ; AR1 points to x(n)
delay *AR1 ; Contents of x(n)is copied to x(n÷1)
These three basic building blocks can be connected to form a block diagram repre-
sentation of a DSP system. The input±output (I/O) description of a DSP system consists
of mathematical expressions with addition, multiplication, and delays, which explicitly
define the relationship between the input and output signals. DSP algorithms are closely
related to block diagram realizations of the I/O difference equations. For example,
consider a simple DSP system described by the difference equation
y(n) = ax(n) ÷ax(n ÷1): (3:1:13)
The block diagram of the system using the three basic building blocks is sketched in
Figure 3.5(a). Note that the difference equation (3.1.13) and the block diagram show
exactly how the output signal y(n) is computed in the DSP system for a given input
signal, x(n).
The DSP algorithm shown in Figure 3.5(a) requires two multiplications and one
addition to compute the output sample y(n). A simple algebraic simplification may
DIGITAL SIGNALS AND SYSTEMS 81
y(n)
Σ
+
+
a
a
(a)
y(n)
Σ
+
+
a
x(n) x(n−1) x(n) x(n−1)
z
−1
z
−1
(b)
Figure 3.5 Block diagrams of DSP systems: (a) direct realization described in (3.1.13), and
(b) simplified implementation given in (3.1.14)
be used to reduce computational requirements. For example, (3.1.13) can be rewritten
as
y(n) = a[x(n) ÷x(n ÷1)[: (3:1:14)
The block diagram implementation of this difference equation is illustrated in Figure
3.5(b), where only one multiplication is required. This example shows that with careful
design (or optimization), the complexity of the system (or algorithm) can be further
reduced.
The C55x implementation of (3.1.14) can be written as:
amov #alpha, XAR1 ; AR1 points to
amov #temp, XAR2 ; AR2 points to temp
mov *(x1n), AC0 ; AC0 = x1(n)
add *(x2n), AC0 ; AC0 = x1(n)÷x2(n)
mov AC0, *AR2 ; Temp = x1(n)÷x2(n), pointed by AR2
mpy *AR1, *AR2, AC1 ; AC1 =
+
[x1(n)÷x2(n)]
Equation (3.1.14) can also be implemented as:
amov #x1n, XAR1 ; AR1 points to x1(n)
amov #x2n, XAR2 ; AR2 points to x2(n)
amov #alpha, XAR3 ; AR3 points to
mpy *AR1, *AR3, AC1 ; AC1 =
+
x1(n)
mac *AR2, *AR3, AC1 ; AC1 =
+
x1(n) ÷
+
x2(n)
When the multiplier coefficient a is a number with a base of 2 such as 0.25 (1/4), we
can use shift operation instead of multiplication. The following example uses the
absolute addressing mode:
mov *(x1n)<#-2, AC0 ; AC0 = 0.25*x1(n)
add *(x2n)<#-2, AC0 ; AC0 = 0.25*x1(n) ÷ 0.25*x2(n)
where the right shift option, <#-2, shifts the content of x1n and x2n to the right by 2
bits (equivalent to dividing it by 4) before they are used.
82 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
3.1.3 Impulse Response of Digital Systems
If the input signal to the DSP system is the unit-impulse sequence d(n) defined in (3.1.1),
then the output signal, h(n), is called the impulse response of the system. The impulse
response plays a very important role in the study of DSP systems. For example, consider
a digital system with the I/O equation
y(n) = b
0
x(n) ÷b
1
x(n ÷1) ÷b
2
x(n ÷2): (3:1:15)
The impulse response of the system can be obtained by applying the unit-impulse
sequence d(n) to the input of the system. The outputs are the impulse response coeffi-
cients computed as follows:
h(0) = y(0) = b
0
1 ÷b
1
0 ÷b
2
0 = b
0
h(1) = y(1) = b
0
0 ÷b
1
1 ÷b
2
0 = b
1
h(2) = y(2) = b
0
0 ÷b
1
0 ÷b
2
1 = b
2
h(3) = y(3) = b
0
0 ÷b
1
0 ÷b
2
0 = 0
. . .
Therefore the impulse response of the system defined in (3.1.15) is ¦b
0
, b
1
, b
2
, 0, 0, . . .¦.
The I/O equation given in (3.1.15) can be generalized as the difference equation with
L parameters, expressed as
y(n) = b
0
x(n) ÷b
1
x(n ÷1) ÷ ÷b
L÷1
x(n ÷L ÷1) =
¸
L÷1
l=0
b
l
x(n ÷l) : (3:1:16)
Substituting x(n) = d(n) into (3.1.16), the output is the impulse response expressed
as
÷1
h(n) =
¸
L÷1
l=0
b
l
d(n ÷l) =
b
n
n = 0 , 1, ..., L
0 otherwise.

(3:1:17)
Therefore the length of the impulse response is L for the difference equation defined in
(3.1.16). Such a system is called a finite impulse response (FIR) system (or filter). The
impulse response coefficients, b
l
, l = 0, 1, . . . , L ÷1, are called filter coefficients
(weights or taps). The FIR filter coefficients are identical to the impulse response
coefficients. Table 3.2 shows the relationship of the FIR filter impulse response h(n)
and its coefficients b
l
.
3.2 Introduction to Digital Filters
As shown in (3.1.17), the system described in (3.1.16) has a finite number of non-zero
impulse response coefficients b
l
, l = 0, 1, . . . , L ÷1. The signal-flow diagram of the
INTRODUCTION TO DIGITAL FILTERS 83
Table 3.2 Relationship of impulse response and coefficients of an FIR filter
b
l
b
0
b
1
b
2
. . . b
L÷1
n x(n) x(n ÷1) x(n ÷2) . . . x(n ÷L ÷1) y(n) = h(n)
0 1 0 0 . . . 0 h(0) = b
0
1 0 1 0 . . . 0 h(1) = b
1
2 0 0 1 . . . 0 h(2) = b
2
. . . . . . . . . . . . . . . . . . . . .
L ÷1 0 0 0 . . . 1 h(L ÷1) = b
L÷1
L 0 0 0 . . . 0 0
y(n)
Σ
+
+
z
−1
z
−1
b
1
b
0
x(n) x(n−1)
+
x(n−L+1)
b
L−1
Figure 3.6 Detailed signal-flow diagram of FIR filter
system described by the I/O Equation (3.1.16) is illustrated in Figure 3.6. The string
of z
÷1
functions is called a tapped-delay-line, as each z
÷1
corresponds to a delay of
one sampling period. The parameter, L, is the order (length) of the FIR filter. The
design and implementation of FIR filters (transversal filters) will be discussed in
Chapter 5.
3.2.1 FIR Filters and Power Estimators
The moving (running) average filter is a simple example of an FIR filter. Averaging is
used whenever data fluctuates and must be smoothed prior to interpretation. Consider
an L-point moving-average filter defined as
y(n) =
1
L
[x(n) ÷x(n ÷1) ÷ ÷x(n ÷L ÷1)[
=
1
L
¸
L÷1
l=0
x(n ÷l), (3:2:1)
where each output signal y(n) is the average of L consecutive input signal samples. The
summation operation that adds all samples of x(n) between 1 and L can be implemented
using the MATLAB statement:
yn = sum(xn(1:L));
84 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
Implementation of (3.2.1) requires L ÷1 additions and L memory locations for
storing signal sequence x(n), x(n ÷1), . . . , x(n ÷L ÷1) in a memory buffer. As illus-
trated in Figure 3.7, the signal samples used to compute the output signal at time n are
L samples included in the window at time n. These samples are almost the same as those
samples used for the previous window at time n ÷1 to compute y(n ÷1), except that the
oldest sample x(n ÷L) of the window at time n ÷1 is replaced by the newest sample
x(n) of the window at time n. Thus (3.2.1) can be computed as
y(n) = y(n ÷1) ÷
1
L
[x(n) ÷x(n ÷L)[: (3:2:2)
Therefore the averaged signal, y(n), can be computed recursively as expressed
in (3.2.2). This recursive equation can be realized by using only two additions.
However, we need L ÷1 memory locations for keeping L ÷1 signal samples
¦x(n)x(n ÷1) x(n ÷L)¦.
The following C5xx assembly code illustrates the implementation of a moving-
average filter of L = 8 based on Equation (3.2.2):
L .set 8 ; Order of filter
xin .usect "indata", 1
xbuffer .usect "indata", L ; Length of buffer
y .usect "outdata", 2,1,1 ; Long-word format
amov #xbuffer÷L÷1, XAR3 ; AR3 points to end of x buffer
amov #xbuffer÷L÷2, XAR2 ; AR2 points to next sample
mov dbl(*(y)), AC1 ; AC1 = y(n÷1)in long format
mov *(xin), AC0 ; AC0 = x(n)
sub *AR3, AC0 ; AC0 = x(n) ÷ x(n±L)
add AC0, #-3, AC1 ; AC1 = y(n±1) ÷ 1/L[x(n)±x(n±L)]
mov AC1, dbl(*(y)) ; y(n) = AC1
rpt # (L÷1) ; Update the tapped-delay-line
mov *AR2÷, *AR3÷ ; X(n±1) = x(n)
mov *(xin), AC0 ; Update the newest sample x(n)
mov AC0, *AR3 ; X(n) = input xin
The strength of a digital signal may be expressed in terms of peak value, energy, and
power. The peak value of deterministic signals is the maximum absolute value of the
signal. That is,
M
x
= max
n
¦[x(n)[¦: (3:2:3)
Window at time n
n−L
Time
n−1
n−L+1 n
Window at time n−1
Figure 3.7 Time windows at current time n and previous time n ÷1
INTRODUCTION TO DIGITAL FILTERS 85
The maximum value of the array xn can be found using the MATLAB function
Mx = max(xn);
The energy of the signal x(n) is defined as
E
x
=
¸
n
[x(n)[
2
: (3:2:4)
The energy of a real-valued x(n) can be calculated by the MATLAB statement:
Ex = sum(abs(xn).^ 2);
Periodic signals and random processes have infinite energy. For such signals, an
appropriate definition of strength is power. The power of signal x(n) is defined as
P
x
= lim
L÷·
1
L
¸
L÷1
n=0
[x(n)[
2
: (3:2:5)
If x(n) is a periodic signal, we have
x(n) = x(n ÷kL), (3:2:6)
where k is an integer and L is the period in samples. Any one period of L samples
completely defines a periodic signal. From Figure 3.7, the power of x(n) can be
computed by
P
x
=
1
L
¸
n
l=n÷L÷1
[x(l)[
2
=
1
L
¸
L÷1
l=0
[x(n ÷l)[
2
: (3:2:7)
For example, a real-valued sinewave of amplitude A defined in (3.1.6) has the power
P
x
= 0:5A
2
.
In most real-time applications, the power estimate of real-valued signals at time n can
be expressed as
^
P
x
(n) =
1
L
¸
L÷1
l=0
x
2
(n ÷l): (3:2:8)
Note that this power estimate uses L samples from the most recent sample at time n
back to the oldest sample at time n ÷L ÷1, as shown in Figure 3.7. Following the
derivation of (3.2.2), we have the recursive power estimator
^
P
x
(n) =
^
P
x
(n ÷1) ÷
1
L
[x
2
(n) ÷x
2
(n ÷L)[: (3:2:9)
86 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
To further simplify the algorithm, we assume L is large enough so that
x
2
(n ÷L) ~
^
P
x
(n ÷1) from a statistical point of view. Thus Equation (3.2.9) can be
further simplified to
^
P
x
(n) ~ 1 ÷
1
L

^
P
x
(n ÷1) ÷
1
L
x
2
(n), (3:2:10)
or
^
P
x
(n) ~ (1 ÷a)
^
P
x
(n ÷1) ÷ax
2
(n), (3:2:11a)
where
a =
1
L
: (3:2:11b)
This is the most effective and widely used recursive algorithm for power estimation
because only three multiplication operations and two memory locations are needed. For
example, (3.2.11a) can be implemented by the C statement
pxn = (1.0-alpha)*pxn ÷ alpha*xn*xn;
where alpha = 1=L as defined in (3.2.11b). This C statement shows that we need three
multiplications and only two memory locations for xn and pxn.
For stationary signals, a larger L (longer window) or smaller a can be used for
obtaining a better average. However, a smaller L (shorter window) should be used
for non-stationary signals for better results. In many real-time applications, the square
of signal x
2
(n) used in (3.2.10) and (3.2.11a) can be replaced with its absolute value [x(n)[
in order to reduce further computation. This efficient power estimator will be further
analyzed in Chapter 4 using the z-transform.
3.2.2 Response of Linear Systems
As discussed in Section 3.1.3, a digital system can be completely characterized by its
impulse response h(n). Consider a digital system illustrated in Figure 3.8, where x(n) is
the input signal and y(n) is the output signal. If the impulse response of the system is
h(n), the output of the system can be expressed as
y(n) = x(n) + h(n) =
¸
·
k=÷·
x(k)h(n ÷k) =
¸
·
k=÷·
h(k)x(n ÷k), (3:2:12)
x(n) y(n) = x(n)∗h(n)
h(n)
Figure 3.8 A simple linear system expressed in time domain
INTRODUCTION TO DIGITAL FILTERS 87
where * denotes the linear convolution operation and the operation defined in (3.2.12) is
called the convolution sum. The input signal, x(n), is convoluted with the impulse
response, h(n), in order to yield the output, y(n). We will discuss the computation of
linear convolution in detail in Chapter 5.
As shown in (3.2.12), the I/O description of a DSP system consists of mathematical
expressions, which define the relationship between the input and output signals. The
exact internal structure of the system is either unknown or ignored. The only way to
interact with the system is by using its input and output terminals as shown in Figure
3.8. The system is assumed to be a `black box'. This block diagram representation is a
very effective way to depict complicated DSP systems.
A digital system is called the causal system if and only if
h(n) = 0, n < 0: (3:2:13)
A causal system is one that does not provide a response prior to input application. For a
causal system, the limits on the summation of the Equation (3.2.12) can be modified to
reflect this restriction as
y(n) =
¸
·
k=0
h(k)x(n ÷k): (3:2:14)
Thus the output signal y(n) of a causal system at time n depends only on present and
past input signals, and does not depend on future input signals.
Consider a causal system that has a finite impulse response of length L. That is,
h(n) =
0, n < 0
b
n
, 0 _ n _ L ÷1
0 n _ L .

(3:2:15)
Substituting this equation into (3.2.14), the output signal can be expressed identically to
the Equation (3.1.16). Therefore the FIR filter output can be calculated as the input
sequence convolutes with the coefficients (or impulse response) of the filter.
3.2.3 IIR Filters
A digital filter can be classified as either an FIR filter or an infinite impulse response
(IIR) filter, depending on whether or not the impulse response of the filter is of
finite or infinite duration. Consider the I/O difference equation of the digital system
expressed as
y(n) = bx(n) ÷ay(n ÷1), (3:2:16)
where each output signal y(n) is dependent on the current input signal x(n) and the
previous output signal y(n ÷1). Assuming that the system is causal, i.e., y(n) = 0 for
n < 0 and let x(n) = d(n). The output signals y(n) are computed as
88 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
y(0) = bx(0) ÷ay(÷1) = b,
y(1) = bx(1) ÷ay(0) = ÷ay(0) = ÷ab,
y(2) = bx(2) ÷ay(1) = ÷ay(1) = a
2
b,
. . .
In general, we have
h(n) = y(n) = (÷1)
n
a
n
b, n = 0, 1, 2, . . . , ·: (3:2:17)
This systemhas infinite impulse response h(n) if the coefficients a and b are non-zero. This
system is called an IIR system (or filter). In theory, we can calculate an IIR filter output
y(n) using either the convolution equation (3.2.14) or the I/Odifference equation (3.2.16).
However, it is not computationally feasible using (3.2.14) for the impulse response h(n)
given in (3.2.17), because we cannot deal with an infinite number of impulse response
coefficients. Therefore we must use an I/O difference equation such as the one defined in
(3.2.16) for computing the IIR filter output in practical applications.
The I/O equation of the IIR system given in (3.2.16) can be generalized with the
difference equation
y(n) = b
0
x(n) ÷b
1
x(n ÷1) ÷ ÷b
L÷1
x(n ÷L ÷1) ÷a
1
y(n ÷1) ÷ ÷a
M
y(n ÷M)
=
¸
L÷1
l=0
b
l
x(n ÷l) ÷
¸
M
m=1
a
m
y(n ÷m): (3:2:18)
This IIR system is represented by a set of feedforward coefficients ¦b
l
, l = 0,
1, . . . , L ÷1¦ and a set of feedback coefficients ¦a
m
, m = 1, 2, . . . , M¦. Since the
outputs are fed back and combined with the weighted inputs, this system is an example
of the general class of feedback systems. Note that when all a
m
are zero, Equation
(3.2.18) is identical to (3.1.16). Therefore an FIR filter is a special case of an IIR filter
without feedback coefficients. An FIR filter is also called a non-recursive filter.
The difference equation of IIR filters given in (3.2.18) can be implemented using the
MATLAB function filter as follows:
yn = filter(b, a, xn);
where the vector b contains feedforward coefficients ¦b
l
, l = 0, 1, . . . , L ÷1¦ and the
vector a contains feedback coefficient ¦a
m
, m = 1, 2, . . . , M¦. The signal vectors, xn
and yn, are the input and output buffers of the system. The FIR filter defined in (3.1.16)
can be implemented using MATLAB as
yn = filter(b, 1, xn);
Assuming that L is large enough so that the oldest sample x(n ÷L) can be approxi-
mated using its average, y(n ÷1). The moving-average filter defined in (3.2.2) can be
simplified as
y(n) = 1 ÷
1
L

y(n ÷1) ÷
1
L
x(n) = (1 ÷a)y(n ÷1) ÷ax(n), (3:2:19)
INTRODUCTION TO DIGITAL FILTERS 89
where a is defined in (3.2.11b). This is a simple first-order IIR filter. Design and
implementation of IIR filters will be further discussed in Chapter 6.
3.3 Introduction to Random Variables
In Section 3.1, we treat signals as deterministic, which are known exactly and repeatable
(such as a sinewave). However, the signals encountered in practice are often random
signals such as speech and interference (noise). These random (stochastic) processes can
be described at best by certain probability concepts. In this section, we will briefly
introduce the concept of probability, followed by random variables and random signal
processing.
3.3.1 Review of Probability and Random Variables
An experiment that has at least two possible outcomes is fundamental to the concept of
probability. The set of all possible outcomes in any given experiment is called the sample
space S. An event A is defined as a subset of the sample space S. The probability of
event A is denoted by P(A). Letting A be any event defined on a sample space S, we have
0 _ P(A) _ 1 (3:3:1)
and
P(S) = 1 (3:3:2)
For example, consider the experiment of rolling of a fair die N times (N ÷ ·), we have
S = ¦1 _ A _ 6¦ and P(A) = 1=6.
A random variable, x, is defined as a function that maps all elements from the sample
space S into points on the real line. A random variable is a number whose value depends
on the outcome of an experiment. Given an experiment defined by a sample space with
elements A, we assign to every A a real number x = x(A) according to certain rules.
Consider the rolling of a fair die Ntimes and assign an integer number to each face of a die,
we have a discrete random variable that can be any one of the discrete values from 1 to 6.
The cumulative probability distribution function of a random variable x is defined as
F(X) = P(x _ X), (3:3:3)
where X is a real number ranging from ÷· to ·, and P(x _ X) is the probability of
¦x _ X¦. Some properties of F(X) are summarized as follows:
F(÷·) = 0, (3:3:4a)
F(·) = 1, (3:3:4b)
0 _ F(X) _ 1, (3:3:4c)
90 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
F(X
1
) _ F(X
2
) if X
1
_ X
2
, (3:3:4d)
P(X
1
< x _ X
2
) = F(X
2
) ÷F(X
1
): (3:3:4e)
The probability density function of a random variable x is defined as
f (X) =
dF(X)
dX
(3:3:5)
if the derivative exists. Some properties of f(X) are summarized as follows:
f (X) _ 0 for all X (3:3:6a)
P(÷· _ x _ ·) =

·
÷·
f (X)dX = 1, (3:3:6b)
F(X) =

X
÷·
f (x)dx, (3:3:6c)
P(X
1
< x _ X
2
) = F(X
2
) ÷F(X
1
) =

X
2
X
1
f (X)dX: (3:3:6d)
Note that both F(X) and f(X) are non-negative functions. The knowledge of these two
functions completely defines the random variable x.
Example 3.2: Consider a random variable x that has a probability density function
f (X) =
0, x < X
1
or x > X
2
a, X
1
_ x _ X
2
,

which is uniformly distributed between X
1
and X
2
. The constant value a can be
computed by using (3.3.6b). That is,

·
÷·
f (X)dX =

X
2
X
1
a dX = a[X
2
÷X
1
[ = 1:
Thus we have
a =
1
X
2
÷X
1
:
If a random variable x is equally likely to take on any value between the two limits X
1
and X
2
and cannot assume any value outside that range, it is said to be uniformly
distributed in the range [X
1
, X
2
]. As shown in Figure 3.9, a uniform density function is
defined as
INTRODUCTION TO RANDOM VARIABLES 91
f (X)
X
0 X
1
X
2
1
X
2
−X
1
Figure 3.9 A uniform density function
f (X) =
1
X
2
÷X
1
, X
1
_ X _ X
2
0, otherwise.

(3:3:7)
This uniform density function will be used to analyze quantization noise in Section 3.4.
If x is a discrete random variable that can take on any one of the discrete values
X
i
, i = 1, 2, . . . as the result of an experiment, we define
p
i
= P(x = X
i
): (3:3:8)
3.3.2 Operations on Random Variables
We can use certain statistics associated with random variables. These statistics
areoften more meaningful from a physical viewpoint than the probability density
function. For example, the mean and the variance are used to reveal sufficient
features of random variables. The mean (expected value) of a random variable x is
defined as
m
x
= E[x[ =

·
÷·
Xf (X)dX, continuous-time case
=
¸
i
X
i
p
i
, discrete-time case, (3:3:9)
where E [] denotes the expectation operation (ensemble averaging).
The expectation is a linear operation. Two useful properties of the expectation
operation are E[a[ = a and E[ax[ = aE[x[, where a is a constant. If E[x[ = 0, x is the
zero-mean random variable. The MATLAB function mean calculates the mean value.
For example, the statement
mx = mean(x);
computes the mean value of the elements in the vector x.
In (3.3.9), the sum is taken over all possible values of x. The mean m
x
defines the
level about which the random process x fluctuates. For example, consider the rolling
of a die N times (N ÷ ·), the probability of outcomes is listed in Table 3.3, as
follows:
92 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
Table 3.3 Probability of rolling a die
X
i
1 2 3 4 5 6
p
i
1/6 1/6 1/6 1/6 1/6 1/6
The mean of outcomes can be computed as
m
x
=
¸
6
i=1
p
i
X
i
=
1
6
(1 ÷2 ÷3 ÷4 ÷5 ÷6) = 3:5:
The variance of x, which is a measure of the spread about the mean, is defined as
s
2
x
= E[(x ÷m
x
)
2
[
=

·
÷·
(X ÷m
x
)
2
f (X)dX, continuous-time case
=
¸
i
p
i
(X
i
÷m
x
)
2
, discrete-time case, (3:3:10)
where x ÷m
x
is the deviation of x from the mean value m
x
. The mean of the squared
deviation indicates the average dispersion of the distribution about the mean m
x
. The
positive square root s
x
of variance is called the standard deviation of x. The MATLAB
function std calculates standard deviation. The statement
s = std(x);
computes the standard deviation of the elements in the vector x.
The variance defined in (3.3.10) can be expressed as
s
2
x
= E[(x ÷m
x
)
2
[ = E[x
2
÷2xm
x
÷m
2
x
[ = E[x
2
[ ÷2m
x
E[x[ ÷m
2
x
= E[x
2
[ ÷m
2
x
:
(3:3:11)
We call E[x
2
[ the mean-square value of x. Thus the variance is the difference between
the mean-square value and the square of the mean value. That is, the variance is the
expected value of the square of the random variable after the mean has been removed.
The expected value of the square of a random variable is equivalent to the notation of
power. If the mean value is equal to 0, then the variance is equal to the mean-square
value. For a zero-mean random variable x, i.e., m
x
= 0, we have
s
2
x
= E[x
2
[ = P
x
, (3:3:12)
which is the power of x. In addition, if y = ax where a is a constant, it can be shown that
s
2
y
= a
2
s
2
x
. It can also be shown (see exercise problem) that P
x
= m
2
x
÷s
2
x
if m
x
= 0.
Consider a uniform density function as given in (3.3.7). The mean of the function can
be computed by
INTRODUCTION TO RANDOM VARIABLES 93
m
x
= E[x[ =

·
÷·
Xf (X)dX =
1
X
2
÷X
1

X
2
X
1
XdX
=
X
2
÷X
1
2
:
(3:3:13)
The variance of the function is
s
2
x
= E[x
2
[ ÷m
2
x
=

·
÷·
X
2
f (X)dX ÷m
2
x
=
1
X
2
÷X
1

X
2
X
1
X
2
dX ÷m
2
x
=
1
X
2
÷X
1
:
X
3
2
÷X
3
1
3
÷m
2
x
=
X
2
2
÷X
1
X
2
÷X
2
1
3
÷
(X
2
÷X
1
)
2
4
=
(X
2
÷X
1
)
2
12
:
(3:3:14)
In general, if x is a uniformly distributed random variable in the interval (÷D, D), the
mean is 0 (m
x
= 0) and the variance is s
2
x
= D
2
=3.
Example 3.3: The MATLAB function rand generates pseudo-random numbers
uniformly distributed in the interval (0, 1). From Equation (3.3.13), the mean of
the generated pseudo-random numbers is 0.5. From (3.3.14), the variance is
computed as 1/12. To generate zero-mean random numbers, we subtract 0.5
from every generated number. The numbers are now distributed in the interval
(÷0.5, 0.5). To make these pseudo-random numbers with unit-variance, i.e.,
s
2
x
= D
2
=3 = 1, the generated numbers must be equally distributed in the interval

3

,

3

). Therefore we have to multiply 2

3

to every generated number that
was subtracted by 0.5. The following MATLAB statement can be used to generate
the uniformly distributed random numbers with mean 0 and variance 1:
xn = 2*sqrt(3)*(rand÷0:5);
For two random variables x and y, we have
E[x ÷y[ = E[x[ ÷E[y[, (3:3:15)
i.e., the mean value of the sum of random variables equals the sum of mean values. The
correlation of x and y is denoted as E[xy]. In general, E[xy[ = E[x[ E[y[. However, if x
and y are uncorrelated, then the correlation can be written in the form
E[xy[ = E[x[E[y[: (3:3:16)
Statistical independence of x and y is sufficient to guarantee that they are uncorrelated.
If the random variables x
i
are independent with the mean m
i
and variance s
2
i
, the
random variable y is defined as
y = x
1
÷x
2
÷ ÷x
N
=
¸
N
i=1
x
i
: (3:3:17)
94 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
m
y
y
f (y)
1/s
y
√2p
Figure 3.10 Probability density function of Gaussian random variable
The probability density function f(Y) becomes a Gaussian (normal) distribution func-
tion (normal curve) as N ÷ ·. That is,
f (Y) =
1
s
y

2p
e
÷(y÷m
y
)
2
=2s
2
y
=
1
s
y

2p
e
÷
1
2
y ÷m
y
s
y

2
¸ ¸
, (3:3:18)
where m
y
=
¸
N
i=1
m
i
and s
y
=

¸
N
i=1
s
2
i

. A graphical representation of the probability
density function defined in (3.3.18) is illustrated in Figure 3.10.
The central limit theorem defined in (3.3.17) is useful in generating a Gaussian
random variable from a uniformly distributed random variable using N _ 12. The
Gaussian random variable is frequently used in communication theory. The MATLAB
function randn generates pseudo-random numbers normally distributed with mean 0
and variance 1.
3.4 Fixed-Point Representation and Arithmetic
The basic element in digital hardware is the two-state (binary) device that contains one
bit of information. A register (or memory unit) containing B bits of information is called
a B-bit word. There are several different methods of representing numbers and carrying
out arithmetic operations. In fixed-point arithmetic, the binary point has a fixed loca-
tion in the register. In floating-point arithmetic, it does not. In general, floating-point
processors are more expensive and slower than fixed-point devices. In this book, we
focus on widely used fixed-point implementations.
A B-bit fixed-point number can be interpreted as either an integer or a fractional
number. It is better to limit the fixed-point representation to fractional numbers because
it is difficult to reduce the number of bits representing an integer. In fixed-point
fractional implementation, it is common to assume that the data is properly scaled so
that their values lie between ÷1 and 1. When multiplying these normalized fractional
numbers, the result (product) will always be less than one.
A given fractional number x has a fixed-point representation as illustrated in
Figure 3.11. In the figure, M is the number of data (magnitude) bits. The most
significant bit
FIXED-POINT REPRESENTATION AND ARITHMETIC 95
x = b
0

.
b
1
b
2

...
b
M–1
b
M
Binary point
Sign-bit
Figure 3.11 Fixed-point representation of binary fractional numbers
b
0
=
0, x _ 0 (positive number)
1, x < 0 (negative number),

(3:4:1)
represents the sign of the number. It is called the sign-bit. The remaining M bits give the
magnitude of the number. The rightmost bit, b
M
, is called the least significant bit (LSB).
The wordlength is B (= M ÷1) bits, i.e., each data point is represented by B ÷1
magnitude bits and one sign-bit.
As shown in Figure 3.11, the decimal value of a positive B-bit binary fractional
number x can be expressed as
(x)
10
= b
1
2
÷1
÷b
2
2
÷2
÷ ÷b
M
2
÷M
=
¸
M
m=1
b
m
2
÷m
: (3:4:2)
For example, the largest (positive) 16-bit fractional number is
x = 0111 1111 1111 1111. The decimal value of this number can be obtained as
x =
¸
15
m=1
2
÷m
= 2
÷1
÷2
÷2
÷ ÷2
÷15
=
¸
15
m=0
1
2

m
÷1 =
1 ÷(1=2)
16
1 ÷(1=2)
÷1
= 1 ÷2
÷15
~ 0:999969:
The negative numbers can be represented using three different formats: the sign-
magnitude, the 1's complement, and the 2's complement. Fixed-point DSP devices
usually use the 2's complement to represent negative numbers because it allows the
processor to perform addition and subtraction uses the same hardware. A positive
number (b
0
= 0) is represented as a simple binary value, while a negative number
(b
0
= 1) is represented using the 2's complement format. With the 2's complement
form, a negative number is obtained by complementing all the bits of the positive binary
number and then adding a 1 to the least significant bit. Table 3.4 shows an example of 3-
bit binary fractional numbers using the 2's complement format and their corresponding
decimal values.
In general, the decimal value of a B-bit binary fractional number can be calculated as
(x)
10
= ÷b
0
÷
¸
15
m=1
b
m
2
÷m
: (3:4:3)
For example, the smallest (negative) 16-bit fractional number is
x = 1000 0000 0000 0000. From (3.4.3), its decimal value is ÷1. Therefore the range of
fractional binary numbers is
96 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
Table 3.4 Example of 3-bit binary fractional numbers in 2's complement format
and their corresponding decimal values
Binary number 000 001 010 011 100 101 110 111
Decimal value 0.00 0.25 0.50 0.75 ÷1.00 ÷0.75 ÷0.50 ÷0.25
÷1 _ x _ (1 ÷2
÷M
): (3:4:4)
For a 16-bit fractional number x, the decimal value range is ÷1 _ x _ 1 ÷2
÷15
.
It is important to note that we use an implied binary point to represent the binary
fractional number. It affects only the accuracy of the result and the location from which
the result will be read. The binary point is purely a programmer's convention and has no
relationship to the hardware. That is, the processor treats the 16-bit number as an
integer. The programmer needs to keep track of the binary point when manipulating
fractional numbers in assembly language programming. For example, if we want to
initialize a data memory location x with the constant decimal value 0.625, we can use
the binary form x = 0101 0000 0000 0000b, the hexadecimal form x = 0x5000, or the
decimal integer x = 2
12
÷2
14
= 20 480. The easiest way to convert a normalized frac-
tional number into the integer that can be used by the C55x assembler is to multiply the
decimal value by 2
15
= 32 768. For example, 0:625 + 32 768 = 20 480.
Most commercially available DSP devices, such as the TMS320C55x discussed in
Chapter 2, are 16-bit processors. These fixed-point DSP devices assume the binary point
after the sign-bit as shown in Figure 3.11. This fractional number representation is also
called the Q15 format since there are 15 magnitude bits.
Example 3.4: The following are some examples of the Q15 format data used for
C55x assembly programming. The directives .set and .equ have the same
functions that assign a value to a symbolic name. They do not require memory
space. The directives .word and .int are used to initialize memory locations
with particular data values represented in binary, hexadecimal, or integer format.
Each data is treated as a 16-bit value and separated by a comma.
ONE .set 32767 ; 1 ÷2
÷15
~ 0:999969 in integer
ONE_HALF .set 0x4000 ; 0.5 in hexadecimal
ONE_EIGHTH .equ 1000h ; 1/8 in hexadecimal
MINUS_ONE .equ 0xffff ; ÷1 in hexadecimal
COEFF .int 0ff00h ; COEFF of ÷0x100
ARRAY .word 2048, ÷2048 ; ARRAY [0.0625, ÷0.0625]
Fixed-point arithmetic is often used with DSP hardware for real-time processing
because it offers fast operation and relatively economical implementation. Its draw-
backs include a small dynamic range (the range of numbers that can be represented) and
low accuracy. Roundoff errors exist only for multiplication. However, the addition may
cause an accumulator overflow. These problems will be discussed in detail in the
following sections.
FIXED-POINT REPRESENTATION AND ARITHMETIC 97
3.5 Quantization Errors
As discussed in Section 3.4, digital signals and system parameters are represented by a
finite number of bits. There is a noticeable error between desired and actual results ± the
finite-precision (finite wordlength, or numerical) effects. In general, finite-precision
effects can be broadly categorized into the following classes:
1. Quantization errors
a. Input quantization
b. Coefficient quantization
2. Arithmetic errors
a. Roundoff (truncation) noise
b. Overflow
The limit cycle oscillation is another phenomenon that may occur when implementing a
feedback system such as an IIR filter with finite-precision arithmetic. The output of the
system may continue to oscillate indefinitely while the input remains 0. This can happen
because of quantization errors or overflow.
This section briefly analyzes finite-precision effects in DSP systems using fixed-point
arithmetic, and presents methods for confining these effects to acceptable levels.
3.5.1 Input Quantization Noise
The ADC shown in Figure 1.2 converts a given analog signal x(t) into digital form x(n).
The input signal is first sampled to obtain the discrete-time signal x(nT). Each x(nT)
value is then encoded using B-bit wordlength to obtain the digital signal x(n), which
consists of M magnitude bits and one sign-bit as shown in Figure 3.11. As discussed in
Section 3.4, we assume that the signal x(n) is scaled such that ÷1 _ x(n) < 1. Thus the
full-scale range of fractional numbers is 2. Since the quantizer employs B bits, the
number of quantization levels available for representing x(nT) is 2
B
. Thus the spacing
between two successive quantization levels is
D =
full-scale range
number of quantization levels
=
2
2
B
= 2
÷B÷1
= 2
÷M
, (3:5:1)
which is called the quantization step (interval, width, or resolution).
Common methods of quantization are rounding and truncation. With rounding, the
signal value is approximated using the nearest quantization level. When truncation is
used, the signal value is assigned to the highest quantization level that is not greater than
the signal itself. Since the truncation produces bias effect (see exercise problem), we use
rounding for quantization in this book. The input value x(nT) is rounded to the nearest
level as illustrated in Figure 3.12. We assume there is a line between two quantization
levels. The signal value above this line will be assigned to the higher quantization level,
while the signal value below this line is assigned to the lower level. For example, the
98 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
000
001
010
011
Quantization level
Time, t
x(t)
0 T 2T
∆ / 2
e(n)

Figure 3.12 Quantization process related to ADC
discrete-time signal x(T) is rounded to 010, since the real value is below the middle line
between 010 and 011, while x(2T) is rounded to 011 since the value is above the middle
line.
The quantization error (noise), e(n), is the difference between the discrete-time signal,
x(nT), and the quantized digital signal, x(n). The error due to quantization can be
expressed as
e(n) = x(n) ÷x(nT): (3:5:2)
Figure 3.12 clearly shows that
[e(n)[ _
D
2
: (3:5:3)
Thus the quantization noise generated by an ADC depends on the quantization interval.
The presence of more bits results in a smaller quantization step, therefore it produces
less quantization noise.
From (3.5.2), we can view the ADC output as being the sum of the quantizer input
x(nT) and the error component e(n). That is,
x(n) = Q[x(nT)[ = x(nT) ÷e(n), (3:5:4)
where Q[] denotes the quantization operation. The nonlinear operation of the quantizer
is modeled as a linear process that introduces an additive noise e(n) to the discrete-time
signal x(nT) as illustrated in Figure 3.13. Note that this model is not accurate for low-
amplitude slowly varying signals.
For an arbitrary signal with fine quantization (B is large), the quantization error e(n)
may be assumed to be uncorrelated with the digital signal x(n), and can be assumed to
be random noise that is uniformly distributed in the interval ÷
D
2
,
D
2

. From (3.3.13), we
can show that
E[e(n)[ =
÷D=2 ÷D=2
2
= 0: (3:5:5)
QUANTIZATION ERRORS 99
x(n)
+
+
Σ
e(n)
x(nT)
Figure 3.13 Linear model for the quantization process
That is, the quantization noise e(n) has zero mean. From (3.3.14) and (3.5.1), we can
show that the variance
s
2
e
=
D
2
12
=
2
÷2B
3
: (3:5:6)
Therefore the larger the wordlength, the smaller the input quantization error.
If the quantization error is regarded as noise, the signal-to-noise ratio (SNR) can be
expressed as
SNR =
s
2
x
s
2
e
= 3 2
2B
s
2
x
, (3:5:7)
where s
2
x
denotes the variance of the signal, x(n). Usually, the SNR is expressed in
decibels (dB) as
SNR = 10 log
10
s
2
x
s
2
e

= 10 log
10
(3 2
2B
s
2
x
)
= 10 log
10
3 ÷20Blog
10
2 ÷10 log
10
s
2
x
= 4:77 ÷6:02B ÷10 log
10
s
2
x

(3:5:8)
This equation indicates that for each additional bit used in the ADC, the converter
provides about 6-dB signal-to-quantization-noise ratio gain. When using a 16-bit ADC
(B = 16), the SNR is about 96 dB. Another important fact of (3.5.8) is that the SNR is
proportional to s
2
x
. Therefore we want to keep the power of signal as large as possible.
This is an important consideration when we discuss scaling issues in Section 3.6.
In digital audio applications, quantization errors arising from low-level signals are
referred to as granulation noise. It can be eliminated using dither (low-level noise) added
to the signal before quantization. However, dithering reduces the SNR. In many applica-
tions, the inherent analog audio components (microphones, amplifiers, or mixers) noise
may already provide enough dithering, so adding additional dithers may not be necessary.
If the digital filter is a linear system, the effect of the input quantization noise alone on
the output may be computed. For example, for the FIR filter defined in (3.1.16), the
variance of the output noise due to the input quantization noise may be expressed as
s
2
y;e
= s
2
e
¸
L÷1
l=0
b
2
l
: (3:5:9)
100 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
This noise is relatively small when compared with other numerical errors and is deter-
mined by the wordlength of ADC.
Example 3.5: Input quantization effects may be subjectively evaluated by observ-
ing and listening to the quantized speech. A speech file called timitl.asc
(included in the software package) was digitized using f
s
= 8 kHz and B = 16.
This speech file can be viewed and played using the MATLAB script:
load(timitl.asc);
plot(timitl);
soundsc(timitl, 8000, 16);
where the MATLAB function soundsc autoscales and plays the vector as sound.
We can simulate the quantization of data with 8-bit wordlength by
qx = round(timitl/256);
where the function, round, rounds the real number to the nearest integer. We then
evaluate the quantization effects by
plot(qx);
soundsc(qx, 8000, 16);
By comparing the graph and sound of timitl and qx, the quantization effects
may be understood.
3.5.2 Coefficient Quantization Noise
When implementing a digital filter, the filter coefficients are quantized to the word-
length of the DSP hardware so that they can be stored in the memory. The filter
coefficients, b
l
and a
m
, of the digital filter defined by (3.2.18) are determined by a filter
design package such as MATLAB for given specifications. These coefficients are usually
represented using the floating-point format and have to be encoded using a finite
number of bits for a given fixed-point processor. Let b
/
l
and a
/
m
denote the quantized
values corresponding to b
l
and a
m
, respectively. The difference equation that can
actually be implemented becomes
y(n) =
¸
L÷1
l=0
b
/
l
x(n ÷l) ÷
¸
M
m=1
a
/
m
y(n ÷m): (3:5:10)
This means that the performance of the digital filter implemented on the DSP hardware
will be slightly different from its design specification. Design and implementation of
digital filters for real-time applications will be discussed in Chapter 5 for FIR filters and
Chapter 6 for IIR filters.
If the wordlength B is not large enough, there will be undesirable effects. The
coefficient quantization effects become more significant when tighter specifications
are used. This generally affects IIR filters more than it affects FIR filters. In many
applications, it is desirable for a pole (or poles) of IIR filters to lie close to the unit circle.
QUANTIZATION ERRORS 101
Coefficient quantization can cause serious problems if the poles of desired filters are too
close to the unit circle because those poles may be shifted on or outside the unit circle
due to coefficient quantization, resulting in an unstable implementation. Such undesir-
able effects due to coefficient quantization are far more pronounced when high-order
systems (where L and M are large) are directly implemented since a change in the value
of a particular coefficient can affect all the poles. If the poles are tightly clustered for a
lowpass or bandpass filter with narrow bandwidth, the poles of the direct-form realiza-
tion are sensitive to coefficient quantization errors. The greater the number of clustered
poles, the greater the sensitivity.
The coefficient quantization noise is also affected by the different structures for the
implementation of digital filters. For example, the direct-form implementation of IIR
filters is more sensitive to coefficient quantization errors than the cascade structure
consisting of sections of first- or second-order IIR filters. This problem will be further
discussed in Chapter 6.
3.5.3 Roundoff Noise
As shown in Figure 3.3 and (3.1.11), we may need to compute the product y(n) = ax(n)
in a DSP system. Assuming the wordlength associated with a and x(n) is B bits, the
multiplication yields 2B bits product y(n). For example, a 16-bit number times another
16-bit number will produce a 32-bit product. In most applications, this product may
have to be stored in memory or output as a B-bit word. The 2B-bit product can be either
truncated or rounded to B bits. Since truncation causes an undesired bias effect, we
should restrict our attention to the rounding case.
In C programming, rounding a real number to an integer number can be implemented
by adding 0.5 to the real number and then truncating the fractional part. For example,
the following C statement
y = (int)(x+0.5);
rounds the real number x to the nearest integer y. As shown in Example 3.5, MATLAB
provides the function round for rounding a real number.
In TMS320C55x implementation, the CPU rounds the operands enclosed by the
rnd() expression qualifier. For example,
mov rnd(HI(AC0)), *AR1
This instruction will round the content of the high portion of AC0(31:16)and the
rounded 16-bit value is stored in the memory location pointed at by AR1. Another key
word, R (or r), when used with the operation code, also performs rounding operation on
the operands. The following is an example that rounds the product of AC0 and AC1
and stores the rounded result in the upper portion of the accumulator AC1(31:16)and
the lower portion of the accumulator AC1(15:0) is cleared:
mpyr AC0, AC1
The process of rounding a 2B-bit product to B bits is very similar to that of quantiz-
ing discrete-time samples using a B-bit quantizer. Similar to (3.5.4), the nonlinear
102 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
operation of product roundoff can be modeled as the linear process shown in Figure
3.13. That is,
y(n) = Q[ax(n)[ = ax(n) ÷e(n), (3:5:11)
where ax(n) is the 2B-bit product and e(n) is the roundoff noise due to rounding 2B-bit
product to B-bit. The roundoff noise is a uniformly distributed random process in the
interval defined in (3.5.3). Thus it has a zero-mean and its power is defined in (3.5.6).
It is important to note that most commercially available fixed-point DSP devices such
as the TMS320C55x have double-precision (2B-bit) accumulator(s). As long as the
program is carefully written, it is quite possible to ensure that rounding occurs only at
the final stage of calculation. For example, consider the computation of FIR filter
output given in (3.1.16). We can keep all the temporary products, b
l
x(n ÷l) for
l = 0, 1, . . . , L ÷1, in the double-precision accumulator. Rounding is only performed
when computation is completed and the sum of products is saved to memory with B-bit
wordlength.
3.6 Overflow and Solutions
Assuming that the input signals and filter coefficients have been properly normalized
(in the range of ÷1 and 1) for fixed-point arithmetic, the addition of these two B-bit
numbers will always produce a B-bit sum. Therefore no roundoff error is introduced by
addition. Unfortunately, when these two B-bit numbers are added, the sum may fall
outside the range of ÷1 and 1. The term overflow is a condition in which the result of
an arithmetic operation exceeds the capacity of the register used to hold that result.
For example, assuming a 3-bit fixed-point hardware with fractional 2's complement
data format (see Table 3.4) is used. If x
1
= 0:75 (011 in binary form) and x
2
= 0:25
(001), the binary sum of x
1
÷x
2
is 100. The decimal value of the binary number 100 is
÷1, not the correct answer ÷1. That is, when the result exceeds the full-scale range of
the register, overflow occurs and unacceptable error is produced. Similarly, subtraction
may result in underflow.
When using a fixed-point processor, the range of numbers must be carefully examined
and adjusted in order to avoid overflow. For the FIR filter defined in (3.1.16), this
overflow results in the severe distortion of the output y(n). For the IIR filter defined in
(3.2.18), the effect is much more serious because the errors are fed back and render the
filter useless. The problem of overflow may be eliminated using saturation arithmetic
and proper scaling (or constraining) signals at each node within the filter to maintain
the magnitude of the signal.
3.6.1 Saturation Arithmetic
Most commercially available DSP devices (such as the TMS320C55x) have mechanisms
that protect against overflow and indicate if it occurs. Saturation arithmetic prevents a
OVERFLOW AND SOLUTIONS 103
x
y
−1
−1
1 − 2
−M
1 − 2
−M
Figure 3.14 Transfer characteristic of saturation arithmetic
result from overflowing by keeping the result at a maximum (or minimum for an
underflow) value. Saturation logic is illustrated in Figure 3.14 and can be expressed as
y =
1 ÷2
÷M
,
x,
÷1,
x _ 1 ÷2
÷M
÷1 _ x < 1
x < ÷1,

(3:6:1)
where x is the original addition result and y is the saturated adder output. If the adder is
under saturation mode, the undesired overflow can be avoided since the 32-bit accu-
mulator fills to its maximum (or minimum) value, but does not roll over. Similar to the
previous example, when 3-bit hardware with saturation arithmetic is used, the addition
result of x
1
÷x
2
is 011, or 0.75 in decimal value. Compared with the correct answer 1,
there is an error of 0.25. However, the result is much better than the hardware without
saturation arithmetic.
Saturation arithmetic has a similar effect to `clipping' the desired waveform. This is a
nonlinear operation that will add undesired nonlinear components into the signal. There-
fore saturation arithmetic can only be used to guarantee that overflow will not occur. It is
not the best, nor the only solution, for solving overflow problems.
3.6.2 Overflow Handling
As mentioned earlier, the TMS320C55x supports the data saturation logic in the data
computation unit (DU) to prevent data computation from overflowing. The logic is
enabled when the overflow mode bit (SATD) in status register ST1 is set (SATD = 1).
When the mode is set, the accumulators are loaded with either the largest positive 32-bit
value (0x00 7FFFFFFF) or the smallest negative 32-bit value (0xFF 8000 0000) if the
result overflows. The overflow mode bit can be set with the instruction
bset SATD
and reset (disabled) with the instruction
bclr SATD
104 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
The TMS320C55x provides overflow flags that indicate whether or not an arithmetic
operation has exceeded the capability of the corresponding register. The overflow flag
ACOVx, (x = 0, 1, 2, or 3) is set to 1 when an overflow occurs in the corresponding
accumulator ACx. The corresponding overflow flag will remain set until a reset is
performed or when a status bit clear instruction is implemented. If a conditional
instruction that tests overflow status (such as a branch, a return, a call, or a conditional
execution) is executed, the overflow flag is cleared. The overflow flags can be tested and
cleared using instructions.
3.6.3 Scaling of Signals
The most effective technique in preventing overflow is by scaling down the magnitudes
of signals at certain nodes in the DSP system and then scaling the result back up to the
original level. For example, consider the simple FIR filter illustrated in Figure 3.15(a).
Let x(n) = 0:8 and x(n ÷1) = 0:6, the filter output y(n) = 1:2. When this filter is
implemented on a fixed-point DSP hardware without saturation arithmetic, undesired
overflow occurs and we get a negative number as a result.
As illustrated in Figure 3.15(b), the scaling factor, b < 1, can be used to scale
down the input signal and prevent overflowing. For example, when b = 0:5 is used,
we have x(n) = 0:4 and x(n ÷1) = 0:3, and the result y(n) = 0:6. This effectively
prevents the hardware overflow. Note that b = 0:5 can be implemented by right shifting
1 bit.
If the signal x(n) is scaled by b, the corresponding signal variance changes to b
2
s
2
x
.
Thus the signal-to-quantization-noise ratio given in (3.5.8) changes to
SNR = 10 log
10
b
2
s
2
x
s
2
e

= 4:77 ÷6:02B ÷10 log
10
s
2
x
÷20 log
10
b: (3:6:2)
Since we perform fractional arithmetic, b < 1 is used to scale down the input signal. The
term 20 log
10
b has negative value. Thus scaling down the amplitude of the signal
reduces the SNR. For example, when b = 0:5, 20 log
10
b = ÷6:02 dB, thus reducing
the SNR of the input signal by about 6 dB. This is equivalent to losing 1-bit in
representing the signal.
y(n)
+
+
0.8 0.9
z
−1
(a)
y(n)
+
+
0.8 0.9
x(n) x(n) x(n−1) x(n−1)
b
z
−1
(b)
Σ Σ
Figure 3.15 Block diagram of simple FIR filters: (a) without scaling, and (b) with scaling
factor b
OVERFLOW AND SOLUTIONS 105
Therefore we have to keep signals as large as possible without overflow. In the FIR
filter given in Figure 3.6, a scaling factor, b, can be applied to the input signal, x(n), to
prevent overflow during the computation of y(n) defined in (3.1.16). The value of signal
y(n) can be bounded as
[y(n)[ = b
¸
L÷1
l=0
b
l
x(n ÷l)

_ bM
x
¸
L÷1
l=0
[b
l
[, (3:6:3)
where M
x
is the peak value of x(n) defined in (3.2.3). Therefore we can ensure that
[y(n)[ < 1 by choosing
b _
1
M
x
¸
L÷1
l=0
[b
l
[
: (3:6:4)
Note that the input signal is bounded and [x(n)[ _ 1, thus M
x
_ 1. The sum
¸
L÷1
l=0
[b
l
[
can be calculated using the MATLAB statement
bsum = sum(abs(b));
where b is the coefficient vector.
Scaling the input by the scaling factor given in (3.6.4) guarantees that overflow never
occurs in the FIR filter. However, the constraint on b is overly conservative for most
signals of practical interest. We can use a more relaxed condition
b _
1
M
x

¸
L÷1
l=0
b
2
l
: (3:6:5)
Other scaling factors that may be used are based on the frequency response of the filter
(will be discussed in Chapter 4). Assuming that the reference signal is narrowband,
overflow can be avoided for all sinusoidal signals if the input is scaled by the maximum
magnitude response of the filter. This scaling factor is perhaps the easiest to use,
especially for IIR filters. It involves calculating the magnitude response and then
selecting its maximum value.
An IIR filter designed by a filter design package such as MATLAB may have some of
its filter coefficients greater than 1.0. To implement a filter with coefficients larger than
1.0, we can also scale the filter coefficients instead of changing the incoming signal. One
common solution is to use a different Q format instead of the Q15 format to represent
the filter coefficients. After the filtering operation is completed, the filter output needs
to be scaled back to the original signal level. This issue will be discussed in the C55x
experiment given in Section 3.8.5.
The TMS320C55x provides four 40-bit accumulators as introduced in Chapter 2.
Each accumulator is split into three parts as illustrated in Figure 3.16. The guard bits
are used as a head-margin for computations. These guard bits prevent overflow in
iterative computations such as the FIR filtering of L _ 256 defined in (3.1.16).
106 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
b39−b32 b31−b16 b15−b0
G H L
Guard bits High-order bits Low-order bits
Figure 3.16 TMS320C55x accumulator configuration
Because of the potential overflow in a fixed-point processor, engineers need to be
concerned with the dynamic range of numbers. This requirement usually demands a
greater coding effort and testing using real data for a given application. In general, the
optimum solution for the overflow problem is by combining scaling factors, guard bits,
and saturation arithmetic. The scaling factors are set as large as possible (close to but
smaller than 1) and occasional overflow can be avoided by using guard bits and
saturation arithmetic.
3.7 Implementation Procedure for Real-Time Applications
The digital filters and algorithms can be implemented on a DSP chip such as the
TMS320C55x following a four-stage procedure to minimize the amount of time spent
on finite wordlength analysis and real-time debugging. Figure 3.17 shows a flowchart of
this procedure.
In the first stage, algorithm design and study is performed on a general-purpose
computer in a non-real-time environment using a high-level MATLAB or C program
with floating-point coefficients and arithmetic. This stage produces an `ideal' system.
In the second stage, we develop the C (or MATLAB) program in a way that emulates
the same sequence of operations that will be implemented on the DSP chip, using the
same parameters and state variables. For example, we can define the data samples and
filter coefficients as 16-bit integers to mimic the wordlength of 16-bit DSP chips. It is
carefully redesigned and restructured, tailoring it to the architecture, the I/O timing
structure, and the memory constraints of the DSP device. This program can also serve
as a detailed outline for the DSP assembly language program or may be compiled using
the DSP chip's C compiler. This stage produces a `practical' system.
The quantization errors due to fixed-point representation and arithmetic can be
evaluated using the simulation technique illustrated in Figure 3.18. The testing data
x(n) is applied to both the ideal system designed in stage 1 and the practical system
developed in stage 2. The output difference, e(n), between these two systems is due to
finite-precision effects. We can re-optimize the structure and algorithm of the practical
system in order to minimize finite-precision errors.
The third stage develops the DSP assembly programs (or mixes C programs with
assembly routines) and tests the programs on a general-purpose computer using a DSP
software simulator (CCS with simulator or EVM) with test data from a disk file. This test
data is either a shortened version of the data used in stage 2, which can be generated
internally by the programor read in as digitized data emulating a real application. Output
from the simulator is saved as another disk file and is compared to the corresponding
output of the C program in the second stage. Once a one-to-one agreement is obtained
between these two outputs, we are assured that the DSP assembly program is essentially
correct.
IMPLEMENTATION PROCEDURE FOR REAL-TIME APPLICATIONS 107
Algorithm analysis
and C program
implementation
Rewrite C program
to emulate
DSP device
DSP assembly program
implementation and
testing by simulator
Real-time testing
in
target system
Start
End
Figure 3.17 Implementation procedure for real-time DSP applications
x(n)
+

e(n)

Practical
system
Ideal
system
Figure 3.18 An efficient technique to study finite wordlength effects
The final stage downloads the compiled (or assembled) and linked program into the
target hardware (such as EVM) and brings it to a real-time operation. Thus the real-
time debugging process is primarily constrained to debugging the I/O timing structure
and testing the long-term stability of the algorithm. Once the algorithm is running, we
can again `tune' the parameters of the systems in a real-time environment.
3.8 Experiments of Fixed-Point Implementations
The purposes of experiments in this section are to learn input quantization effects and to
determine the proper fixed-point representation for a DSP system.
108 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
3.8.1 Experiment 3A ± Quantization of Sinusoidal Signals
To experiment with input quantization effects, we shift off (right) bits of input signal
and then evaluate the shifted samples. Altering the number of bits for shifting right, we
can obtain an output stream that corresponds to a wordlength of 14 bits, 12 bits, and so
on. The example given in Table 3.5 simulates an A/D converter of different wordlength.
Instead of shifting the samples, we mask out the least significant 4 (or 8, or 10) bits of
each sample, resulting in the 12 (8 or 6) bits data having comparable amplitude to the
16-bit data.
1. Copy the C function exp3a.c and the linker command file exp3.cmd from the
software package to A: \Experiment3 directory, create project exp3a to simulate
16, 12, 8, and 6 bits A/D converters. Use the run-time support library rts55.lib
and build the project.
2. Use the CCS graphic display function to plot all four output buffers: out16,
out12, out8, and out6. Examples of the plots and graphic settings are shown in
Figure 3.19 and Figure 3.20, respectively.
3. Compare the graphic results of each output stream, and describe the differences
between waveforms represented by different wordlength.
Table 3.5 Program listing of quantizing a sinusoid, exp3a.c
#define BUF_SIZE 40
const int sineTable [BUF_SIZE] =
{0x0000, 0x01E0, 0x03C0, 0x05A0, 0x0740, 0x08C0, 0x0A00, 0x0B20,
0x0BE0, 0x0C40, 0x0C60, 0x0C40, 0x0BE0, 0x0B20, 0x0A00, 0x08C0,
0x0740, 0x05A0, 0x03C0, 0x01E0, 0x0000, 0xFE20, 0xFC40, 0xFA60,
0xF8C0, 0xF740, 0xF600, 0xF4E0, 0xF420, 0xF3C0, 0xF3A0, 0xF3C0,
0xF420, 0xF4E0, 0xF600, 0xF740, 0xF8C0, 0xFA60, 0xFC40, 0x0000};
int out16 [BUF_SIZE]; /* 16 bits output sample buffer */
int out12 [BUF_SIZE]; /* 12 bits output sample buffer */
int out8 [BUF_SIZE]; /* 8 bits output sample buffer */
int out6 [BUF_SIZE]; /* 6 bits output sample buffer */
void main()
{
int i;
for (i = 0; i < BUF_SIZE÷1; i÷÷)
{
out16[i]= sineTable[i]; /* 16-bit data */
out12[i]= sineTable[i]&0xfff0; /* Mask off 4-bit */
out8[i]= sineTable[i]&0xff00; /* Mask off 8-bit */
out6[i]= sineTable[i]&0xfc00; /* Mask off 10-bit */
}
}
EXPERIMENTS OF FIXED-POINT IMPLEMENTATIONS 109
Figure 3.19 Quantizing 16-bit data (top-left) into 12-bit (bottom-left), 8-bit (top-right), and
6-bit (bottom-right)
Figure 3.20 Example of displaying graphic settings
110 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
4. Find the mean and variance of quantization noise for the 12-, 8-, and 6-bit A/D
converters.
3.8.2 Experiment 3B ± Quantization of Speech Signals
There are many practical applications from cellular phones to MP3 players that process
speech (audio) signals using DSP. To understand the quantization effects of speech
signals, we use a digitized speech file, timit1.asc, as the input for this experiment. An
experiment code for this experiment, exp3b.c, is listed in Table 3.6.
1. Refer to the program listed in Table 3.6. Write a C function called exp3b.c to
simulate 16, 12, 8, and 4 bits A/D converters (or copy the file exp3b.c from the
software package). Use the digitized speech file timit1.asc (included in the soft-
ware package) as the input signal for the experiment. Create the project exp3b, add
exp3b.c and exp3.cmd into the project.
2. Use CCS probe points to connect disk files as described in Chapter 1. In this experi-
ment, we use a probe point toconnect the input speechtothe variable named indata.
We also connect four output variables out16, out12, out8, and out4 to generate
quantized output files. As mentioned in Chapter 1, we need to add a header line to the
input data file, timit1.asc. The header information is formatted in the following
syntax using hexadecimal numbers:
Magic Number Format Starting Address Page Number Length
1651 2 C4 1 1
where
. the magic number is always set to 1651
Table 3.6 Program listing of quantizing a speech signal
#define LENGTH 27956 /* Length of input file timitl.asc */
int indata, out16, out12, out8, out4;
void main( )
{
int i;
for(i = 0; i < LENGTH; i÷÷)
out16 = indata; /* Simulate 16-bit A/D */
out12 = indata&0xfff0; /* Simulate 12-bit A/D */
out8 = indata&0xff00; /* Simulate 8-bit A/D */
out4 = indata&0xf000; /* Simulate 4-bit A/D */
}
EXPERIMENTS OF FIXED-POINT IMPLEMENTATIONS 111
. the format is 2 for ASCII data file format
. the starting address is the address of the data variable where we want to connect
the input file to, such as C4 in the above example
. the page number is 1 for data
. the length defines the number of data to pass each time
3. Invoke the CCS and set probe points for input and outputs in exp3b.c. Use
probe points to output the speech signals with wordlength of 16, 12, 8, and 4 bits
to data files. Because the output file generated by CCS probe points have a
header line, we need to remove this header. If we want to use MATLAB to listen
to the output files, we have to set the CCS output file in integer format. We can
load the output file out12.dat and listen to it using the following MATLAB
commands:
load out12.dat; % Read data file
soundsc(out12, 8000, 16); % Play at 8kHz
Listen to the quantization effects between the files with different wordlength.
3.8.3 Experiment 3C ± Overflow and Saturation Arithmetic
As discussed in Section 3.6, overflow may occur when DSP algorithms perform accumu-
lations such as FIR filtering. When the number exceeds the maximum capacity of an
accumulator, overflow occurs. Sometimes an overflow occurs when data is transferred to
memory even though the accumulator does not overflow. This is because the C55x
accumulators (AC0±AC3) have 40 bits, while the memory space is usually defined as a
16-bit word. There are several ways to handle the overflow. As introduced in Section 3.6,
the C55x has a built-in overflow-protection unit that will saturate the data value if
overflow occurs.
In this experiment, we will use an assembly routine, ovf_sat.asm (included in
the software package), to evaluate the results with and without overflow protection.
Table 3.7 lists a portion of ovf_sat.asm.
In the program, the following code repeatedly adds the constant 0x140 to AC0:
rptblocal add_loop_end÷1
add #0x140<#16, AC0
mov hi(AC0), *AR5÷
add_loop_end
The updated value is stored at the buffer pointed at by AR5. The content of AC0 will
grow larger and larger and eventually the accumulator will overflow. When the over-
flow occurs, a positive number in AC0 suddenly becomes negative. However, when the
C55x saturation mode is enabled, the overflowed positive number will be limited to
0x7FFFFFFF.
112 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
Table 3.7 List of assembly program to observe overflow and saturation
.def _ovftest
.bss buff, (0x100)
.bss buff1, (0x100)
;
; Code start
;
_ovftest
bclr SATD ; Clear saturation bit if set
xcc start, T0 != #0 ; If T0 != 0, set saturation bit
bset SATD
start
pshboth XAR5 ; Save XAR5
... ... ; Some instructions omitted here
mov #0, AC0
mov #0x80÷1, BRC0 ; Initialize loop counts for addition
amov #buff÷0x80, XAR5 ; Initialize buffer pointer
rptblocal add_loop_end÷1
add #0x140<#16, AC0 ; Use AC0 as a ramp up counter
mov hi(AC0), *AR5÷ ; Save the counter to buffer
add_loop_end
... ... ; Some instructions omitted here
mov #0x100÷1, BRC0 ; Init loop counts for sinewave
amov #buff1, XAR5 ; Initialize buffer pointer
mov mmap(@AR0), BSA01 ; Initialize base register
mov #40, BK03 ; Set buffer size to 40
mov #20, AR0 ; Start with an offset of 20
bset AR0LC ; Activate circular buffer
rptblocal sine_loop_end÷1
mov *ar0÷ <#16, AC0 ; Get sine value into high AC0
sfts AC0, #9 ; Scale the sine value
mov hi(AC0), *AR5÷ ; Save scaled value
sine_loop_end
mov #0, T0 ; Return 0 if no overflow
xcc set_ovf_flag, overflow(AC0)
mov #1, T0 ; Return 1 if overflow detected
set_ovf_flag
bclr AR0LC ; Reset circular buffer bit
bclr SATD ; Reset saturation bit
popboth XAR5 ; Restore AR5
ret
The second portion of the code stores the left-shifted sinewave values to data memory
locations. Without saturation protection, this shift will cause some of the shifted values
to overflow.
EXPERIMENTS OF FIXED-POINT IMPLEMENTATIONS 113
As we mentioned in Chapter 2, circular addressing is a very useful addressing mode.
The following segment of code in the example shows how to set up and use the circular
addressing mode:
mov #sineTable, BSA01 ; Initialize base register
mov #40, BK03 ; Set buffer size to 40
mov #20, AR0 ; Start with an offset of 20
bset AR0LC ; Activate circular buffer
The first instruction sets up the circular buffer base register (BSAxx). The second
instruction initializes the size of the circular buffer. The third instruction initializes the
offset from the base as the starting point. In this case, the offset is set to 20 words from
the base sineTable[]. The last instruction enables AR0 as the circular pointer. The
program exp3c.c included in the software package is used to call the assembly routine
ovf_sat.asm, and is shown in Figure 3.21.
1. Create the project exp3c that uses exp3c.c and ovf_sat.asm (included in the
software package) for this experiment.
2. Use the graphic function to display the sinewave (top) and the ramp counter
(bottom) as shown in Figure 3.22.
Figure 3.21 Experiment of C program exp3c.c
114 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
(a) (b)
Figure 3.22 C55x data saturation example: (a) without saturation protection, and (b) with
saturation protection enabled
3.8.4 Experiment 3D ± Quantization of Coefficients
Filters are widely used for DSP applications. The C55x implementation of FIR (or IIR)
filters often use 16-bit numbers to represent filter coefficients. Due to the quantization
of coefficients, the filter implemented using fixed-point hardware will not have the exact
same response as the filter that is obtained by a filter design package, such as
MATLAB, which uses the floating-point numbers to represent coefficients. Since filter
design will be discussed in Chapters 5 and 6, we only briefly describe the fourth-order
IIR filter used in this experiment.
Table 3.8 shows an assembly language program that implements a fourth-order IIR
lowpass filter. This filter is designed for 8 kHz sampling frequency with cut-off fre-
quency 1800 Hz. The routine, _init_iir4, initializes the memory locations of x and y
buffers to 0. The IIR filter routine, _iir4, filters the input signal. The coefficient data
pointer (CDP) is used to point to the filter coefficients. The auxiliary registers, AR5 and
AR6, are pointing to the x and y data buffers, respectively. After each sample is
processed, both the x and y buffers are updated by shifting the data in the tapped-
delay-line.
1. Write a C function, exp3d.c, to call the routine _ii4() to perform lowpass filter
operation. The initialization needs to be done only once, while the routine _ii4()
will be called for filtering every sample, these files are also included in the software
package.
2. The filter coefficient quantization effects can be observed by modifying the MASK
value defined in assembly routine _ii4(). Adjusting the MASK to generate 12, 10,
and 8 bits quantized coefficients. Interface the C function, exp3d.c with a signal
source and use either the simulator (or EVM) to observe the quantization effects due
to the limited wordlength representing filter coefficients.
EXPERIMENTS OF FIXED-POINT IMPLEMENTATIONS 115
Table 3.8 List of C55x assembly program for a fourth-order IIR filter
; exp3d_IIR.asm ÷ fourth-order IIR filter
;
MASK .set 0xFFFF
.def _iir4
.def _init_iir4
;
; Original coefficients of fourth-order IIR LPF
; with sampling frequency of 8000Hz
;
; int b[5]= {0.0072, 0.00287, 0.0431, 0.0287, 0.0072};
; int a[5]= {1.0000, ÷2.16860, 2.0097, ÷0.8766, 0.1505};
;
.data; ;Q13 formatted coefficients
coeff ; b0, b1, b2, b3, b4
.word 0x003B&MASK, 0x00EB&MASK
.word 0x0161&MASK, 0x00EB&MASK, 0x003B&MASK
; ÷a1, ÷a2, ÷a3, ÷a4
.word 0x4564&MASK, ÷0x404F&MASK
.word 0xC0D&MASK, ÷0x04D1&MASK
.bss x, 5 ; x delay line
.bss y, 4 ; y delay line
.text
_init_iir4
pshboth XAR5
amov #x, XAR5
rpt #4
mov #0, *AR5÷
amov #y, XAR5
rpt #3
mov #0, *AR5÷
popboth XAR5
ret
;
; Fourth-order IIR filter
; Entry T0 = sample
; Exit T0 = filtered sample
;
_iir4
pshboth XAR5
pshboth XAR6
bset SATD
bset SXM
amov #x, XAR5
amov #y, XAR6
116 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
Table 3.8 (continued )
amov #coeff, XCDP
bset FRCT
[ [ mov T0, *AR5 ; x[0]= indata
;
; Perform IIR filtering
;
mpym *AR5÷, *CDP÷, AC0 ; AC0 = x[0]*bn[0]
[ [ rpt #3 ; i = 1, 2, 3, 4
macm *AR5÷, *CDP÷, AC0 ; AC0 ÷= x[i]*bn[i]
rpt #3 ; i = 0, 1, 2, 3
macm *AR6÷, *CDP÷, AC0 ; AC0 ÷= y[i]*an[i]
amov #y÷2, XAR5
amov #y÷3, XAR6
sfts AC0, #2 ; Scale to Q15 format
[ [ rpt #2
mov *AR5÷, *AR6÷ ; Update y []
mov hi(AC0), *AR6
[ [ mov hi(AC0), T0 ; Return y[0]in T0
amov #x÷3, XAR5
amov #x÷4, XAR6
bclr FRCT
[ [ rpt #3
mov *AR5÷, *AR6÷ ; Update x[]
popboth XAR6
popboth XAR5
bclr SXM
bclr SATD
[ [ ret
.end
3.8.5 Experiment 3E ± Synthesizing Sine Function
For many DSP applications, signals and system parameters, such as filter coefficients,
are usually normalized in the range of ÷1 and 1 using fractional numbers. In Section
3.4, we introduced the fixed-point representation of fractional numbers and in Section
3.6, we discussed the overflow problems and present some solutions. The experiment in
this section used polynomial approximation of the sinusoid function as an example to
understand the fixed-point arithmetic operation and overflow control.
The cosine and sine functions can be expressed as the infinite power (Taylor) series
expansion
cos(y) = 1 ÷
1
2!
y
2
÷
1
4!
y
4
÷
1
6!
y
6
÷ , (3:8:1a)
EXPERIMENTS OF FIXED-POINT IMPLEMENTATIONS 117
sin(y) = y ÷
1
3!
y
3
÷
1
5!
y
5
÷
1
7!
y
7
÷ , (3:8:1b)
where y is in radians and `!' represents the factorial operation. The accuracy of the
approximation depends on how many terms are used in the series. Usually more terms
are needed to provide reasonable accuracy for larger values of y. However, in real-time
DSP application, only a limited number of terms can be used. Using a function
approximation approach such as the Chebyshev approximation, cos(y) and sin(y) can
be computed as
cos(y) =1 ÷0:001922y ÷4:9001474y
2
÷0:264892y
3
÷ 5:04541y
4
÷1:800293y
5
,
(3:8:2a)
sin(y) = 3:140625y ÷0:02026367y
2
÷5:325196y
3
÷0:5446788y
4
÷1:800293y
5
,
(3:8:2b)
where the value y is defined in the first quadrant. That is, 0 _ y < p=2. For y in the
other quadrants, the following properties can be used to transfer it to the first quadrant:
sin(180
o
÷y) = sin(y), cos(180
o
÷y) = ÷cos(y), (3:8:3)
sin(÷180
o
÷y) = ÷sin(y), cos(÷180
o
÷y) = ÷cos(y) (3:8:4)
and
sin(÷y) = ÷sin(y), cos(÷y) = cos(y): (3:8:5)
The C55x assembly routine given in Table 3.9 synthesizes the sine and cosine functions,
which can be used to calculate the angle y from ÷180
o
to 180
o
.
As shown in Figure 3.11, data in the Q15 format is within the range defined in (3.4.4).
Since the absolute value of the largest coefficient given in this experiment is 5.325196, we
cannot use the Q15 format to represent this number. To properly represent the coeffi-
cients, we have to scale the coefficient, or use a different Q format that represents both
the fractional numbers and the integers. We can achieve this by assuming the binary
point to be three bits further to the right. This is called the Q12 format, which has one
sign-bit, three integer bits, and 12 fraction bits, as illustrated in Figure 3.23(a). The
Q12 format covers the range ÷8 to 8. In the given example, we use the Q12 format
for all the coefficients, and map the angle ÷p _ y _ p to a signed 16-bit number
(0x8000 _ x _ 0x7FFF), as shown in Figure 3.23(b).
When the sine_cos subroutine is called, a 16-bit mapped angle (function
argument) is passed to the assembly routine in register T0 following the C calling
conversion described in Chapter 2. The quadrant information is tested and stored in
TC1 and TC2. If TC1 (bit 14) is set, the angle is located in either quadrant II or IV. We
use the 2's complement to convert the angle to the first or third quadrant. We mask out
the sign-bit to calculate the third quadrant angle in the first quadrant, and the negation
changes the fourth quadrant angle to the first quadrant. Therefore the angle to be
118 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
Table 3.9 Sine function approximation routine for the C55x
; sine_cos: 16-bit sine(x) and cos(x) approximation function
;
; Entry: T0 = x, in the range of [-pi = 0x8000 pi = 0x7fff]
; AR0 -> Pointer
; Return: None
; Update:
; ARO -> [0]= cos(x) = d0 ÷ d1*x ÷ d2*x^2 ÷ d3*x^3 ÷ d4*x^4 ÷ d5*x^5
; ARO -> [1]= sin(x) = cl*x ÷ c2*x^2 ÷ c3*x^3 ÷ c4*x^4 ÷ c5*x^5
;
.def _sine_cos
;
; Approximation coefficients in Q12 (4096) format
;
.data
coeff ; Sine approximation coefficients
.word 0x3240 ; cl = 3.140625
.word 0x0053 ; c2 = 0.02026367
.word 0xaacc ; c3 = ÷5.325196
.word 0x08b7 ; c4 = 0.54467780
.word 0x1cce ; c5 = 1.80029300
; Cosine approximation coefficients
.word 0x1000 ; d0 = 1.0000
.word 0xfff8 ; d1 = ÷0.001922133
.word 0xb199 ; d2 = ÷4.90014738
.word 0xfbc3 ; d3 = ÷0.2648921
.word 0x50ba ; d4 = 5.0454103
.word 0xe332 ; d5 = ÷1.800293
;
; Function starts
;
.text
_sine_cos
pshboth XAR5 ; Save AR5
amov #14, AR5
btstp AR5, T0 ; Test bit 15 and 14
;
; Start cos(x)
;
amov #coeff÷10, XAR5 ; Pointer to the end of coefficients
xcc _neg_x, TC1
neg T0 ; Negate if bit 14 is set
_neg_x
and #0x7fff, T0 ; Mask out sign bit
mov *AR5÷<#16, AC0 ; AC0 = d5
continues overleaf
EXPERIMENTS OF FIXED-POINT IMPLEMENTATIONS 119
Table 3.9 (continued )
[ [ bset SATD ; Set Saturate bit
mov *AR5÷<#16, AC1 ; AC1 = d4
[ [ bset FRCT ; Set up fractional bit
mac AC0, T0, AC1 ; AC1 = (d5*x ÷ c4)
[ [ mov *AR5÷<#16,AC0 ; AC0 = d3
mac AC1,T0, AC0 ; AC0 = (d5*x^2 ÷ d4*x ÷ d3)
[ [ mov *AR5÷<#16, AC1 ; AC1 = d2
mac AC0, T0, AC1 ; AC1 = (d5*x^3 ÷ d4*x^2 ÷ d3*x ÷ d2)
[ [ mov *AR5÷<#16, AC0 ; AC0 = d1
mac AC1, T0, AC0 ; AC0 = (d5*x^4 ÷ d4*x^3 ÷ d3*x^2
; ÷ d2*x ÷ d1)
[ [ mov *AR5÷<#16, AC1 ; AC1 = d0
macr AC0, T0, AC1 ; AC1 = (d5*x^4 ÷ d4*x^3 ÷ d3*x^2 ÷ d2*x
; ÷ d1)*x ÷ d0
[ [ xcc _neg_result1, TC2
neg AC1
_neg_result1
mov *AR5÷<#16, AC0 ; AC0 = c5
[ [ xcc _neg_result2, TC1
neg AC1
_neg_result2
mov hi(saturate(AC1<#3)), *AR0÷ ; Return cos(x) in Q15
;
; Start sin(x) computation
;
mov *AR5÷<#16, AC1 ; AC1 = c4
mac AC0, T0, AC1 ; AC1 = (c5*x ÷ c4)
[ [ mov *AR5÷<#16, AC0 ; AC0 = c3
mac AC1, T0, AC0 ; AC0 = (c5*x^2 ÷ c4*x ÷ c3)
[ [ mov *AR5÷<#16, AC1 ; AC1 = c2
mac AC0, T0, AC1 ; AC1 = (c5*x^3 ÷ c4*x^2 ÷ c3*x ÷ c2)
[ [ mov *AR5÷<#16, AC0 ; AC0 = c1
mac AC1, T0, AC0 ; AC0 = (c5*x^4 ÷ c4*x^3 ÷ c3*x^2
; ÷ c2*x ÷ c1)
[ [ popboth XAR5 ; Restore AR5
mpyr T0, AC0, AC1 ; AC1 = (c5*x^4 ÷ c4*x^3 ÷ c3*x^2
; ÷ c2*x ÷ c1)*x
[ [ xcc _neg_result3, TC2
neg AC1
_neg_result3
mov hi(saturate(AC1<#3)), *AR0÷ ; Return sin(x)in Q15
[ [ bclr FRCT ; Reset fractional bit
bclr SATD ; Reset saturate bit
ret
.end
120 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
s.xxxxxxxxxxxxxxx
siii.xxxxxxxxxxxx
Q15 format
Q12 format
0xFFFF = 360Њ
0x0000 = 0Њ
0x3FFF = 90Њ
0x7FFF = 180Њ
0x8000 = −180Њ
0xBFFF = −90Њ
(a) (b)
Figure 3.23 Scaled fixed-point number representation: (a) Q formats, and (b) Map angle value
to 16-bit signed integer
calculated is always located in the first quadrant. Because we use the Q12 format
coefficients, the computed result needs to be left shifted 3 bits to become the Q15
format.
1. Write a C function exp3e.c to call the sinusoid approximation function,
sine_cos, written in assembly listed in Table 3.9 (These programs can be found
in the software package). Calculate the angles in the following table.
y 308 458 608 908 1208 1358 1508 1808
cos(y)
sin(y)
y ÷1508 ÷1358 ÷1208 ÷908 ÷608 ÷458 ÷308 3608
cos(y)
sin(y)
2. In the above implementation of sine approximation, what will the following C55x
assembly instructions do? What may happen to the approximation result if we do
not set these control bits?
(a) .bset FRCT
(b) .bset SATD
(c) .bclr FRCT
(d) .bclr SATD
References
[1] N. Ahmed and T. Natarajan, Discrete-Time Signals and Systems, Englewood Cliffs, NJ: Prentice-
Hall, 1983.
[2] MATLAB User's Guide, Math Works, 1992.
REFERENCES 121
[3] MATLAB Reference Guide, Math Works, 1992.
[4] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, Englewood Cliffs, NJ:
Prentice-Hall, 1989.
[5] S. J. Orfanidis, Introduction to Signal Processing, Englewood Cliffs, NJ: Prentice-Hall, 1996.
[6] J. G. Proakis and D. G. Manolakis, Digital Signal Processing ± Principles, Algorithms, and
Applications, 3rd Ed., Englewood Cliffs, NJ: Prentice-Hall, 1996.
[7] P. Peebles, Probability, Random Variables, and Random Signal Principles, New York, NY:
McGraw-Hill, 1980.
[8] A Bateman and W. Yates, Digital Signal Processing Design, New York: Computer Science Press,
1989.
[9] S. M. Kuo and D. R. Morgan, Active Noise Control Systems ± Algorithms and DSP Implementa-
tions, New York: Wiley, 1996.
[10] C. Marven and G. Ewers, A Simple Approach to Digital Signal Processing, New York: Wiley, 1996.
[11] J. H. McClellan, R. W. Schafer, and M. A. Yoder, DSP First: A Multimedia Approach, 2nd Ed.,
Englewood Cliffs, NJ: Prentice-Hall, 1998.
[12] D. Grover and J. R. Deller, Digital Signal Processing and the Microcontroller, Upper Saddle River,
NJ: Prentice-Hall, 1999.
Exercises
Part A
1. Compute the impulse response h(n) for n _ 0 of the digital systems defined by the following
I/O difference equations:
(a) y(n) = x(n) ÷0:75y(n ÷1)
(b) y(n) ÷0:5y(n ÷1) = 2x(n) ÷x(n ÷1)
(c) y(n) = 2x(n) ÷0:75x(n ÷1) ÷1:5x(n ÷2)
2. Construct detailed flow diagrams for the three digital systems defined in Problem 1.
3. Similar to the signal flow diagram for the FIR filter shown Figure 3.6, construct a detailed
signal flow diagram for the IIR filter defined in (3.2.18).
4. A discrete-time system is time invariant (or shift invariant) if its input±output characteristics
do not change with time. Otherwise this system is time varying. A digital system with input
signal x(n) is time invariant if and only if the output signal
y(n ÷k) = F[x(n ÷k)[
for any time shift k, i.e., when an input is delayed (shifted) by k, the output is delayed by the
same amount. Show that the system defined in (3.2.18) is time-invariant system if the
coefficients a
m
and b
l
are constant.
5. A linear system is a system that satisfies the superposition principle, which states that the
response of the system to a weighted sum of signals is equal to the corresponding weighted
sum of the responses of the system to each of the individual input signals. That is, a system is
linear if and only if
122 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
F[a
1
x
1
(n) ÷a
2
x
2
(n)[ = a
1
F[x
1
(n)[ ÷a
2
F[x
2
(n)[
for any arbitrary input signals x
1
(n) and x
2
(n), and for any arbitrary constants a
1
and a
2
. If
the input is the sum (superposition) of two or more scaled sequences, we can find the output
due to each sequence acting alone and then add (superimpose) the separate scaled outputs.
Check whether the following systems are linear or nonlinear:
(a) y(n) = 0:5x(n) ÷0:75y(n ÷1)
(b) y(n) = x(n)x(n ÷1) ÷0:5y(n ÷1)
(c) y(n) = 0:75x(n) ÷x(n)y(n ÷1)
(d) y(n) = 0:5x(n) ÷0:25x
2
(n)
6. Show that a real-valued sinewave of amplitude A defined in (3.1.6) has the power
P
x
= 0:5A
2
.
7. Equation (3.3.12) shows that the power is equal to the variance for a zero-mean random
variable. Show that if the mean of the random variable x is m
x
, the power of x is given by
P
x
= m
2
x
÷s
2
x
.
8. An exponential random variable is defined by the probability density function
f (x) =
l
2
e
÷l[x[
:
Show that the mean value is 0 and the variance is 2=l
2
.
9. Find the fixed-point 2's complement representation with B = 8 for the decimal numbers
0.152 and ÷0.738. Round the binary numbers to 6 bits and compute the corresponding
roundoff errors.
10. If the quantization process uses truncation instead of rounding, show that the truncation
error, e(n) = x(n) ÷x(nT), will be in the interval ÷D < e(n) < 0. Assuming that the trunca-
tion error is uniformly distributed in the interval (÷D, 0), compute the mean and the variance
of e(n).
11. Identify the various types of finite wordlength effects that can occur in a digital filter defined
by the I/O equation (3.2.18).
12. Consider the causal system with I/O equation
y(n) = x(n) ÷0:5y(n ÷1)
and the input signal given as
x(n) =
0:5, n = 0
0, n > 0.

(a) Compute y(0), y(1), y(2), y(3), and y(·).
EXERCISES 123
(b) Assume the DSP hardware has 4-bit worldlength (B = 4), compute y(n) for
n = 0, 1, 2, 3, . . . , ·. In this case, show that y(n) oscillates between ±0:125 for n _ 2.
(c) Repeat part (b) but use wordlength B = 5. Show that the output y(n) oscillates between
±0:0625 for n _ 3.
Part B
13. Generate and plot (20 samples) the following sinusoidal signals using MATLAB:
(a) A = 1, f = 100 Hz, and f
s
= 1 kHz
(b) A = 1, f = 400 Hz, and f
s
= 1 kHz
(c) Discuss the difference of results between (a) and (b)
(d) A = 1, f = 600 Hz, and f
s
= 1 kHz
(e) Compare and explain the results (b) and (d).
14. Generate 1024 samples of pseudo-random numbers with mean 0 and variance 1 using the
MATLAB function rand. Then use MATLAB functions mean, std, and hist to verify the
results.
15. Generate 1024 samples of sinusoidal signal at frequency 1000 Hz, amplitude equal to unity,
and the sampling rate is 8000 Hz. Mix the generated sinewave with the zero-mean pseudo-
random number of variance 0.2.
16. Write a C program to implement the moving-average filter defined in (3.2.2). Test the filter
using the corrupted sinewave generated in Problem 15 as input for different L. Plot both the
input and output waveforms. Discuss the results related to the filter order L.
17. Given the difference equations in Problem 1, calculate and plot the impulse response
h(n), n = 0, 1, . . . , 127 using MATLAB.
18. Assuming that
^
P
x
(0) = 1, use MATLAB to estimate the power of x(n) generated in Problem
15 by using the recursive power estimator given in (3.2.11). Plot
^
P
x
(n) for n = 0, 1, . . . , 1023.
Part C
19. Using EVM (or other DSP boards) to conduct the quantization experiment in real-time:
(a) Generate an analog input signal, such as a sinewave, using a signal generator. Both the
input and output channels of the DSP are displayed on an oscilloscope. Assuming the
ADC has 16-bit resolution and adjusting the amplitude of input signal to the full scale of
ADC without clipping the waveform. Vary the number of bits (by shifting out or
masking) to 14, 12, 10, etc. to represent the signal and output the signal to DAC.
Observe the output waveform using the oscilloscope.
(b) Replace the input source with a microphone, radio line output, or CD player, and send
DSP output to a loudspeaker for audio play back. Vary the number of bits (16, 12, 8, 4,
124 DSP FUNDAMENTALS AND IMPLEMENTATION CONSIDERATIONS
etc.) for the output signal, and listen to the output sound. Depending upon the type of
loudspeaker being used, we may need to use an amplifier to drive the loudspeaker.
20. Implement the following square-root approximation equation in C55x assembly language:

x

= 0:2075806 ÷1:454895x ÷1:34491x
2
÷1:106812x
3
÷0:536499x
4
÷0:1121216x
5
This equation approximates an input variable within the range of 0:5 _ x _ 1. Based on the
values in the following table, calculate

x

.
x 0.5 0.6 0.7 0.8 0.9

x

21. Write a C55x assembly function to implement the inverse square-root approximation
equation as following:
1=

x

= 1:84293985 ÷2:57658958x ÷2:11866164x
2
÷0:67824984x
3
:
This equation approximates an input variable in the range of 0:5 _ x _ 1. Use this approxi-
mation equation to compute 1=

x

in the following table:
x 0.5 0.6 0.7 0.8 0.9
1=

x

Note that 1=

x

will result in a number greater than 1.0. Try to use Q14 data format. That is,
use 0x3FFF for 1 and 0x2000 for 0.5, and scale back to Q15 after calculation.
EXERCISES 125
4
Frequency Analysis
Frequency analysis of any given signal involves the transformation of a time-domain
signal into its frequency components. The need for describing a signal in the frequency
domain exists because signal processing is generally accomplished using systems that are
described in terms of frequency response. Converting the time-domain signals and
systems into the frequency domain is extremely helpful in understanding the character-
istics of both signals and systems.
In Section 4.1, the Fourier series and Fourier transform will be introduced. The
Fourier series is an effective technique for handling periodic functions. It provides a
method for expressing a periodic function as the linear combination of sinusoidal
functions. The Fourier transform is needed to develop the concept of frequency-domain
signal processing. Section 4.2 introduces the z-transform, its important properties, and
its inverse transform. Section 4.3 shows the analysis and implementation of digital
systems using the z-transform. Basic concepts of discrete Fourier transforms will be
introduced in Section 4.4, but detailed treatments will be presented in Chapter 7. The
application of frequency analysis techniques using MATLAB to design notch filters and
analyze room acoustics will be presented in Section 4.5. Finally, real-time experiments
using the TMS320C55x will be presented in Section 4.6.
4.1 Fourier Series and Transform
In this section, we will introduce the representation of analog periodic signals using
Fourier series. We will then expand the analysis to the Fourier transform representation
of broad classes of finite energy signals.
4.1.1 Fourier Series
Any periodic signal, x(t), can be represented as the sum of an infinite number of
harmonically related sinusoids and complex exponentials. The basic mathematical
representation of periodic signal x(t) with period T
0
(in seconds) is the Fourier series
defined as
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
x(t) =
¸
·
k=÷·
c
k
e
jkO
0
t
, (4:1:1)
where c
k
is the Fourier series coefficient, and V
0
= 2p=T
0
is the fundamental frequency
(in radians per second). The Fourier series describes a periodic signal in terms of infinite
sinusoids. The sinusoidal component of frequency kV
0
is known as the kth harmonic.
The kth Fourier coefficient, c
k
, is expressed as
c
k
=
1
T
0

T
0
x(t)e
÷jkV
0
t
dt: (4:1:2)
This integral can be evaluated over any interval of length T
0
. For an odd function, it is
easier to integrate from 0 to T
0
. For an even function, integration from ÷T
0
=2 to T
0
=2
is commonly used. The term with k = 0 is referred to as the DC component because
c
0
=
1
T
0
¸
T
0
x(t)dt equals the average value of x(t) over one period.
Example 4.1: The waveform of a rectangular pulse train shown in Figure 4.1 is a
periodic signal with period T
0
, and can be expressed as
x(t) =
A, kT
0
÷t=2 _ t _ kT
0
÷t=2
0, otherwise,

(4:1:3)
where k = 0, ±1, ±2, F F F , and t < T
0
. Since x(t) is an even signal, it is con-
venient to select the integration from ÷T
0
=2 to T
0
=2. From (4.1.2), we have
c
k
=
1
T
0

T
0
2
÷
T
0
2
Ae
÷jkV
0
t
dt =
A
T
0
e
÷jkV
0
t
÷jkV
0

t
2
÷
t
2
¸ ¸
=
At
T
0
sin
kV
0
t
2

kV
0
t
2
: (4:1:4)
This equation shows that c
k
has a maximum value At=T
0
at V
0
= 0, decays to 0 as
V
0
÷±·, and equals 0 at frequencies that are multiples of p. Because the
periodic signal x(t) is an even function, the Fourier coefficients c
k
are real values.
For the rectangular pulse train with a fixed period T
0
, the effect of decreasing t is to
spread the signal power over the frequency range. On the other hand, when t is fixed but
the period T
0
increases, the spacing between adjacent spectral lines decreases.
t
x(t)
A

t
2
t
2
0
2 2
−T
0
−T
0
T
0
T
0
Figure 4.1 Rectangular pulse train
128 FREQUENCY ANALYSIS
A periodic signal has infinite energy and finite power, which is defined by Parseval's
theorem as
P
x
=
1
T
0

T
0

x(t)

2
dt =
¸
·
k=÷·

c
k

2
: (4:1:5)
Since c
k
[ [
2
represents the power of the kth harmonic component of the signal, the total
power of the periodic signal is simply the sum of the powers of all harmonics.
The complex-valued Fourier coefficients, c
k
, can be expressed as
c
k
= c
k
[ [e
jf
k
: (4:1:6)
A plot of [c
k
[ versus the frequency index k is called the amplitude (magnitude) spectrum,
and a plot of f
k
versus k is called the phase spectrum. If the periodic signal x(t) is real
valued, it is easy to show that c
0
is real valued and that c
k
and c
÷k
are complex
conjugates. That is,
c
k
= c
+
÷k
, c
÷k
[ [ = c
k
[ [ and f
÷k
= ÷f
k
: (4:1:7)
Therefore the amplitude spectrum is an even function of frequency V, and the phase
spectrum is an odd function of V for a real-valued periodic signal.
If we plot [c
k
[
2
as a function of the discrete frequencies kV
0
, we can show that the
power of the periodic signal is distributed among the various frequency components.
This plot is called the power density spectrum of the periodic signal x(t). Since the power
in a periodic signal exists only at discrete values of frequencies kV
0
, the signal has a line
spectrum. The spacing between two consecutive spectral lines is equal to the funda-
mental frequency V
0
.
Example 4.2: Consider the output of an ideal oscillator as the perfect sinewave
expressed as
x(t) = sin 2pf
0
t ( ), f
0
=
V
0
2p
:
We can then calculate the Fourier series coefficients using Euler's formula
(Appendix A.3) as
sin(2pf
0
t) =
1
2j
e
j2pf
0
t
÷e
÷j2pf
0
t

=
¸
·
k=÷·
c
k
e
jk2pf
0
t
:
We have
c
k
=
1=2j, k = 1
÷1=2j, k = ÷1
0, otherwise.

(4:1:8)
FOURIER SERIES AND TRANSFORM 129
This equation indicates that there is no power in any of the harmonic k = ±1.
Therefore Fourier series analysis is a useful tool for determining the quality
(purity) of a sinusoidal signal.
4.1.2 Fourier Transform
We have shown that a periodic signal has a line spectrum and that the spacing between
two consecutive spectral lines is equal to the fundamental frequency V
0
= 2p=T
0
. The
number of frequency components increases as T
0
is increased, whereas the envelope of
the magnitude of the spectral components remains the same. If we increase the period
without limit (i.e., T
0
÷·), the line spacing tends toward 0. The discrete frequency
components converge into a continuum of frequency components whose magnitudes
have the same shape as the envelope of the discrete spectra. In other words, when the
period T
0
approaches infinity, the pulse train shown in Figure 4.1 reduces to a single
pulse, which is no longer periodic. Thus the signal becomes non-periodic and its
spectrum becomes continuous.
In real applications, most signals such as speech signals are not periodic. Consider the
signal that is not periodic (V
0
÷0 or T
0
÷·), the number of exponential components
in (4.1.1) tends toward infinity and the summation becomes integration over the entire
continuous range (÷·, ·). Thus (4.1.1) can be rewritten as
x(t) =
1
2p

·
÷·
X(V)e
jVt
dV: (4:1:9)
This integral is called the inverse Fourier transform. Similarly, (4.1.2) can be rewritten
as
X(V) =

·
÷·
x(t)e
÷jVt
dt, (4:1:10)
which is called the Fourier transform (FT) of x(t). Note that the time functions
are represented using lowercase letters, and the corresponding frequency functions are
denoted by using capital letters. A sufficient condition for a function x(t) that possesses
a Fourier transform is

·
÷·
[x(t)[dt < ·: (4:1:11)
That is, x(t) is absolutely integrable.
Example 4.3: Calculate the Fourier transform of the function x(t) = e
÷at
u(t), where
a > 0 and u(t) is the unit step function. From (4.1.10), we have
130 FREQUENCY ANALYSIS
X(V) =

·
÷·
e
÷at
u(t)e
÷jVt
dt
=

·
0
e
÷(a÷jV)t
dt
=
1
a ÷jV
:
The Fourier transform X(V) is also called the spectrum of the analog signal x(t). The
spectrum X(V) is a complex-valued function of frequency V, and can be expressed as
X(V) =

X(V)

e
jf(V)
, (4:1:12)
where [X(V)[ is the magnitude spectrum of x(t), and f(V) is the phase spectrum of x(t).
In the frequency domain, [X(V)[
2
reveals the distribution of energy with respect to the
frequency and is called the energy density spectrum of the signal. When x(t) is any finite
energy signal, its energy is
E
x
=

·
÷·
[x(t)[
2
dt =
1
2p

·
÷·
[X(V)[
2
dV: (4:1:13)
This is called Parseval's theorem for finite energy signals, which expresses the principle
of conservation of energy in time and frequency domains.
For a function x(t) defined over a finite interval T
0
, i.e., x(t) = 0 for [t[ > T
0
=2, the
Fourier series coefficients c
k
canbe expressedinterms of X(V) using (4.1.2) and(4.1.10) as
c
k
=
1
T
0
X kV
0
( ): (4:1:14)
For a given finite interval function, its Fourier transform at a set of equally spaced
points on the V-axis is specified exactly by the Fourier series coefficients. The distance
between adjacent points on the V-axis is 2p=T
0
radians.
If x(t) is a real-valued signal, we can show from (4.1.9) and (4.1.10) that
FT x(÷t)[ = X
+
(V) and X(÷V) = X
+
(V): [ (4:1:15)
It follows that
[X(÷V)[ = [X(V)[ and f(÷V) = ÷f(V): (4:1:16)
Therefore the amplitude spectrum [X(V)[ is an even function of V, and the phase
spectrum is an odd function.
If the time signal x(t) is a delta function d(t), its Fourier transform can be calculated as
X(V) =

·
÷·
d(t)e
÷jVt
dt = 1: (4:1:17)
FOURIER SERIES AND TRANSFORM 131
This indicates that the delta function has frequency components at all frequencies. In
fact, the narrower the time waveform, the greater the range of frequencies where the
signal has significant frequency components.
Some useful functions and their Fourier transforms are summarized in Table 4.1. We
may find the Fourier transforms of other functions using the Fourier transform proper-
ties listed in Table 4.2.
Table 4.1 Common Fourier transform pairs
Time function x(t) Fourier transform X(V)
d(t) 1
d(t ÷t) e
÷jVt
1 2pd(V)
e
÷at
u(t)
1
a ÷jV
e
jV
0
t
2pd(V÷V
0
)
sin(V
0
t) jp[d(V÷V
0
) ÷d(V÷V
0
)[
cos(V
0
t) p[d(V÷V
0
) ÷d(V÷V
0
)[
sgn(t) =
1, t _ 0
÷1, t < 0

2
jV
Table 4.2 Useful properties of the Fourier transform
Time function x(t) Property Fourier transform X(V)
a
1
x
1
(t) ÷a
2
x
2
(t) Linearity a
1
X
1
(V) ÷a
2
X
2
(V)
dx(t)
dt
Differentiation in time
domain
jVX(V)
tx(t) Differentiation in
frequency domain
j
dX(V)
dV
x(÷t) Time reversal X(÷V)
x(t ÷a) Time shifting e
÷jVa
X(V)
x(at) Time scaling
1
[a[
X
V
a

x(t) sin(V
0
t) Modulation
1
2j
[X(V÷V
0
) ÷X(V÷V
0
)[
x(t) cos(V
0
t) Modulation
1
2
[X(V÷V
0
) ÷X(V÷V
0
)[
e
÷at
x(t) Frequency shifting X(V÷a)
132 FREQUENCY ANALYSIS
Example 4.4: Find the Fourier transform of the time function
y(t) = e
÷a[t[
, a > 0:
This equation can be written as
y(t) = x(÷t) ÷x(t),
where
x(t) = e
÷at
u(t), a > 0:
From Table 4.1, we have X(V) = 1=(a ÷jV). From Table 4.2, we have
Y(V) = X(÷V) ÷X(V). This results in
Y(V) =
1
a ÷jV
÷
1
a ÷jV
=
2a
a
2
÷V
2
:
4.2 The z-Transform
Continuous-time signals and systems are commonly analyzed using the Fourier trans-
form and the Laplace transform (will be introduced in Chapter 6). For discrete-time
systems, the transform corresponding to the Laplace transform is the z-transform. The
z-transform yields a frequency-domain description of discrete-time signals and systems,
and provides a powerful tool in the design and implementation of digital filters. In this
section, we will introduce the z-transform, discuss some important properties, and show
its importance in the analysis of linear time-invariant (LTI) systems.
4.2.1 Definitions and Basic Properties
The z-transform (ZT) of a digital signal, x(n), ÷·< n < ·, is defined as the power
series
X(z) =
¸
·
n=÷·
x(n)z
÷n
, (4:2:1)
where X(z) represents the z-transform of x(n). The variable z is a complex variable, and
can be expressed in polar form as
z = re
jy
, (4:2:2)
where r is the magnitude (radius) of z, and y is the angle of z. When r = 1, [z[ = 1 is
called the unit circle on the z-plane. Since the z-transform involves an infinite power
series, it exists only for those values of z where the power series defined in (4.2.1)
THE Z-TRANSFORM 133
converges. The region on the complex z-plane in which the power series converges is
called the region of convergence (ROC).
As discussed in Section 3.1, the signal x(n) encountered in most practical applications
is causal. For this type of signal, the two-sided z-transform defined in (4.2.1) becomes a
one-sided z-transform expressed as
X(z) =
¸
·
n=0
x(n)z
÷n
: (4:2:3)
Clearly if x(n) is causal, the one-sided and two-sided z-transforms are equivalent.
Example 4.5: Consider the exponential function
x(n) = a
n
u(n):
The z-transform can be computed as
X(z) =
¸
·
n=÷·
a
n
z
÷n
u(n) =
¸
·
n=0
(az
÷1
)
n
:
Using the infinite geometric series given in Appendix A.2, we have
X(z) =
1
1 ÷az
÷1
if [az
÷1
[ < 1:
The equivalent condition for convergence (or ROC) is
[z[ > [a[:
Thus we obtain X(z) as
X(z) =
z
z ÷a
, [z[ > [a[:
There is a zero at the origin z = 0 and a pole at z = a. The ROC and the pole±zero
plot are illustrated in Figure 4.2 for 0 < a < 1, where `×' marks the position of the
pole and `o' denotes the position of the zero. The ROC is the region outside
the circle with radius a. Therefore the ROC is always bounded by a circle since the
convergence condition is on the magnitude [z[. A causal signal is characterized by
an ROC that is outside the maximum pole circle and does not contain any pole.
The properties of the z-transform are extremely useful for the analysis of discrete-time
LTI systems. These properties are summarized as follows:
1. Linearity (superposition). The z-transform is a linear transformation. Therefore the
z-transform of the sum of two sequences is the sum of the z-transforms of the
individual sequences. That is,
134 FREQUENCY ANALYSIS
|z| = a
|z| = 1
Re[z]
Im[z]
Figure 4.2 Pole, zero, and ROC (shaded area) on the z-plane
ZT[a
1
x
1
(n) ÷a
2
x
2
(n)[ = a
1
ZT[x
1
(n)[ ÷a
2
ZT[x
2
(n)[
= a
1
X
1
(z) ÷a
2
X
2
(z), (4:2:4)
2. where a
1
and a
2
are constants, and X
1
(z) and X
2
(z) are the z-transforms of the
signals x
1
(n) and x
2
(n), respectively. This linearity property can be generalized for
an arbitrary number of signals.
2. Time shifting. The z-transform of the shifted (delayed) signal y(n) = x(n ÷k) is
Y(z) = ZT[x(n ÷k)[ = z
÷k
X(z), (4:2:5)
2. where the minus sign corresponds to a delay of k samples. This delay property states
that the effect of delaying a signal by k samples is equivalent to multiplying its
z-transform by a factor of z
÷k
. For example, ZT[x(n ÷1)[ = z
÷1
X(z). Thus the unit
delay z
÷1
in the z-domain corresponds to a time shift of one sampling period in the
time domain.
3. Convolution. Consider the signal
x(n) = x
1
(n) + x
2
(n), (4:2:6)
2. where + denotes the linear convolution introduced in Chapter 3, we have
X(z) = X
1
(z)X
2
(z): (4:2:7)
2. Therefore the z-transform converts the convolution of two time-domain signals to
the multiplication of their corresponding z-transforms.
Some of the commonly used signals and their z-transforms are summarized in
Table 4.3.
THE Z-TRANSFORM 135
Table 4.3 Some common z-transform pairs
x(n), n _ 0, c is constant X(z)
c
cz
z ÷1
cn
cz
(z ÷1)
2
c
n
z
z ÷c
nc
n
cz
(z ÷c)
2
ce
÷an
cz
z ÷e
÷a
sin(!
0
n)
z sin(!
0
)
z
2
÷2z cos(!
0
) ÷1
cos(!
0
n)
z[z ÷cos(!
0
)[
z
2
÷2z cos(!
0
) ÷1
4.2.2 Inverse z-transform
The inverse z-transform can be expressed as
x(n) = ZT
÷1
[X(z)[ =
1
2pj

C
X(z)z
n÷1
dz, (4:2:8)
where C denotes the closed contour in the ROC of X(z) taken in a counterclockwise
direction. Several methods are available for finding the inverse z-transform. We will
discuss the three most commonly used methods ± long division, partial-fraction expan-
sion, and residue method.
Given the z-transform X(z) of a causal sequence, it can be expanded into an infinite
series in z
÷1
or z by long division. To use the long-division method, we express X(z) as
the ratio of two polynomials such as
X(z) =
B(z)
A(z)
=
¸
L÷1
l=0
b
l
z
÷l
¸
M
m=0
a
m
z
÷m
, (4:2:9)
where A(z) and B(z) are expressed in either descending powers of z, or ascending powers
of z
÷1
. Dividing B(z) by A(z) obtains a series of negative powers of z if a positive-time
sequence is indicated by the ROC. If a negative-time function is indicated, we express
X(z) as a series of positive powers of z. The method will not work for a sequence defined
136 FREQUENCY ANALYSIS
in both positive and negative time. In addition, it is difficult to obtain a closed-form
solution of the time-domain signal x(n) via the long-division method.
The long-division method can be performed recursively. That is,
x(n) = b
n
÷
¸
n
m=1
x(n ÷m)a
m
¸ ¸
a
0
, n = 1, 2, F F F (4:2:10)
where
x(0) = b
0
=a
0
: (4:2:11)
This recursive equation can be implemented on a computer to obtain x(n).
Example 4.6: Given
X(z) =
1 ÷2z
÷1
÷z
÷2
1 ÷z
÷1
÷0:3561z
÷2
using the recursive equation given in (4.2.10), we have
x(0) = b
0
=a
0
= 1,
x(1) = [b
1
÷x(0)a
1
[=a
0
= 3,
x(2) = [b
2
÷x(1)a
1
÷x(0)a
2
[=a
0
= 3:6439,
F F F
This yields the time domain signal x(n) = ¦1, 3, 3:6439, F F F¦ obtained from long
division.
The partial-fraction-expansion method factors the denominator of X(z) if it is
not already in factored form, then expands X(z) into a sum of simple partial fractions.
The inverse z-transform of each partial fraction is obtained from the z-transform
tables such as Table 4.3, and then added to give the overall inverse z-transform. In
many practical cases, the z-transform is given as a ratio of polynomials in z or z
÷1
as
shown in (4.2.9). If the poles of X(z) are of first order and M = L ÷1, then X(z) can be
expanded as
X(z) = c
0
÷
¸
L÷1
l=1
c
l
1 ÷p
l
z
÷1
= c
0
÷
¸
L÷1
l=1
c
l
z
z ÷p
l
, (4:2:12)
where p
l
are the distinct poles of X(z) and c
l
are the partial-fraction coefficients. The
coefficient c
l
associated with the pole p
l
may be obtained with
c
l
=
X(z)
z
(z ÷p
l
)

z=p
l
: (4:2:13)
THE Z-TRANSFORM 137
If the order of the numerator B(z) is less than that of the denominator A(z) in (4.2.9),
that is L ÷1 < M, then c
0
= 0. If L ÷1 > M, then X(z) must be reduced first in order to
make L ÷1 _ M by long division with the numerator and denominator polynomials
written in descending power of z
÷1
.
Example 4.7: For the z-transform
X(z) =
z
÷1
1 ÷0:25z
÷1
÷0:375z
÷2
we can first express X(z) in positive powers of z, expressed as
X(z) =
z
z
2
÷0:25z ÷0:375
=
z
(z ÷0:75)(z ÷0:5)
=
c
1
z
z ÷0:75
÷
c
2
z
z ÷0:5
:
The two coefficients are obtained by (4.2.13) as follows:
c
1
=
X(z)
z
(z ÷0:75)

z=0:75
=
1
z ÷0:5

z=0:75
= 0:8
and
c
2
=
1
z ÷0:75

z=÷0:5
= ÷0:8:
Thus we have
X(z) =
0:8z
z ÷0:75
÷
0:8z
z ÷0:5
:
The overall inverse z-transform x(n) is the sum of the two inverse z-transforms.
From entry 3 of Table 4.3, we obtain
x(n) = 0:8[(0:75)
n
÷(÷0:5)
n
[ , n _ 0:
The MATLAB function residuez finds the residues, poles and direct terms of the
partial-fraction expansion of B(z)=A(z) given in (4.2.9). Assuming that the numerator
and denominator polynomials are in ascending powers of z
÷1
, the function
[c, p, g]= residuez(b, a);
finds the partial-fraction expansion coefficients, c
l
, and the poles, p
l
, in the returned
vectors c and p, respectively. The vector g contains the direct (or polynomial) terms of
the rational function in z
÷1
if L ÷1 _ M. The vectors b and a represent the coefficients
of polynomials B(z) and A(z), respectively.
If X(z) contains one or more multiple-order poles, the partial-fraction expansion must
include extra terms of the form
¸
m
j=1
g
j
(z÷p
l
)
j
for an mth order pole at z = p
l
. The
coefficients g
j
may be obtained with
138 FREQUENCY ANALYSIS
g
j
=
1
(m÷j)3
d
m÷j
dz
m÷j
(z ÷p
l
)
m
X(z)
z
¸
z=p
l
: (4:2:14)
Example 4.8: Consider the function
X(z) =
z
2
÷z
(z ÷1)
2
:
We first express X(z) as
X(z) =
g
1
(z ÷1)
÷
g
2
(z ÷1)
2
:
From (4.2.14), we have
g
1
=
d
dz
(z ÷1)
2
X(z)
z
¸ ¸
z=1
=
d
dz
(z ÷1)

z=1
= 1,
g
2
=
(z ÷1)
2
X(z)
z

z=1
= z ÷1) ( [
z=1
= 2:
Thus
X(z) =
z
(z ÷1)
÷
2z
(z ÷1)
2
:
From Table 4.3, we obtain
x(n) = ZT
÷1
z
z ÷1

÷ZT
÷1
2z
(z ÷1)
2
¸ ¸
= 1 ÷2n, n _ 0:
The residue method is based on Cauchy's integral theorem expressed as
1
2pj

c
z
k÷m÷1
dz =
1 if k = m
0 if k = m.

(4:2:15)
Thus the inversion integral in (4.2.8) can be easily evaluated using Cauchy's residue
theorem expressed as
x(n) =
1
2pj

c
X(z)z
n÷1
dz
=
¸
residues of X(z)z
n÷1
at poles of X(z)z
n÷1
within C:
(4:2:16)
THE Z-TRANSFORM 139
The residue of X(z)z
n÷1
at a given pole at z = p
l
can be calculated using the formula
R
z=p
l
=
d
m÷1
dz
m÷1
(z ÷p
l
)
m
(m÷1)3
X(z)z
n÷1
¸
z=p
l
, m _ 1, (4:2:17)
where m is the order of the pole at z = p
l
. For a simple pole, Equation (4.2.17) reduces to
R
z=p
l
= (z ÷p
l
)X z)z
n÷1

z=p
l
: (4:2:18)
Example 4.9: Given the following z-transform function:
X(z) =
1
(z ÷1)(z ÷0:5)
,
we have
X(z)z
n÷1
=
z
n÷1
(z ÷1)(z ÷0:5)
:
This function has a simple pole at z = 0 when n = 0, and no pole at z = 0 for
n _ 1. For the case n = 0,
X(z)z
n÷1
=
1
z(z ÷1)(z ÷0:5)
:
The residue theorem gives
x(n) = R
z=0
÷R
z=1
÷R
z=0:5
= zX z)z
n÷1

z=0
÷(z ÷1)X(z)z
n÷1

z=1
÷ z ÷0:5)X(z)z
n÷1

z=0:5
= 2 ÷2 ÷(÷4) = 0:
For the case that n _ 1, the residue theorem is applied to obtain
x(n) = R
z=1
÷R
z=0:5
= (z ÷1)X(z)z
n÷1

z=1
÷(z ÷0:5)X(z)z
n÷1

z=0:5
= 2 ÷2(0:5)
n÷1
= 2 1 ÷(0:5)
n÷1

, n _ 1:
We have discussed three methods for obtaining the inverse z-transform. A limitation
of the long-division method is that it does not lead to a closed-form solution. However,
it is simple and lends itself to software implementation. Because of its recursive nature,
care should be taken to minimize possible accumulation of numerical errors when the
number of data points in the inverse z-transform is large. Both the partial-fraction-
expansion and the residue methods lead to closed-form solutions. The main disadvan-
tage with both methods is the need to factor the denominator polynomial, which is done
by finding the poles of X(z). If the order of X(z) is high, finding the poles of X(z) may be
140 FREQUENCY ANALYSIS
a difficult task. Both methods may also involve high-order differentiation if X(z)
contains multiple-order poles. The partial-fraction-expansion method is useful in gen-
erating the coefficients of parallel structures for digital filters. Another application of z-
transforms and inverse z-transforms is to solve linear difference equations with constant
coefficients.
4.3 Systems Concepts
As mentioned earlier, the z-transform is a powerful tool in analyzing digital systems. In
this section, we introduce several techniques for describing and characterizing digital
systems.
4.3.1 Transfer Functions
Consider the discrete-time LTI system illustrated in Figure 3.8. The system output is
computed by the convolution sum defined as y(n) = x(n) + h(n). Using the convolution
property and letting ZT[x(n)[ = X(z) and ZT[ y(n)[ = Y(z), we have
Y(z) = X(z)H(z), (4:3:1)
where H(z) = ZT[h(n)[ is the z-transform of the impulse response of the system. The
frequency-domain representation of LTI system is illustrated in Figure 4.3.
The transfer (system) function H(z) of an LTI system may be expressed in terms of the
system's input and output. From (4.3.1), we have
H(z) =
Y(z)
X(z)
= ZT[h(n)[ =
¸
·
n=÷·
h(n)z
÷n
: (4:3:2)
Therefore the transfer function of the LTI system is the rational function of two
polynomials Y(z) and X(z). If the input x(n) is the unit impulse d(n), the z-transform
of such an input is unity (i.e., X(z) = 1), and the corresponding output Y(z) = H(z).
One of the main applications of the z-transform in filter design is that the z-transform
can be used in creating alternative filters that have exactly the same input±output
behavior. An important example is the cascade or parallel connection of two or more
x(n) y(n) = x(n)∗h(n) h(n)
Y(z) = X(z)H(z) X(z) H(z)
ZT
−1
ZT ZT
Figure 4.3 A block diagram of LTI system in both time-domain and z-domain
SYSTEMS CONCEPTS 141
systems, as illustrated in Figure 4.4. In the cascade (series) interconnection, the output
of the first system, y
1
(n), is the input of the second system, and the output of the second
system, y(n), is the overall system output. From Figure 4.4(a), we have
Y
1
(z) = X(z)H
1
(z) and Y(z) = Y
1
(z)H
2
(z):
Thus
Y(z) = X(z)H
1
(z)H
2
(z):
Therefore the overall transfer function of the cascade of the two systems is
H(z) =
Y(z)
X(z)
= H
1
(z)H
2
(z): (4:3:3)
Since multiplication is commutative, H
1
(z)H
2
(z) = H
2
(z)H
1
(z), the two systems can be
cascaded in either order to obtain the same overall system response. The overall impulse
response of the system is
h(n) = h
1
(n) + h
2
(n) = h
2
(n) + h
1
(n): (4:3:4)
Similarly, the overall impulse response and the transfer function of the parallel
connection of two LTI systems shown in Figure 4.4(b) are given by
h(n) = h
1
(n) ÷h
2
(n) (4:3:5)
and
H(z) = H
1
(z) ÷H
2
(z): (4:3:6)
x(n)
x(n)
H
1
(z)
H
1
(z)
H(z)
H(z)
H
2
(z)
H
2
(z)
y
1
(n)
X(z)
y(n)
y(n)
Y(z) Y
1
(z)
y
1
(n)
y
2
(n)
(a)
(b)
Figure 4.4 Interconnect of digital systems: (a) cascade form, and (b) parallel form
142 FREQUENCY ANALYSIS
If we can multiply several z-transforms to get a higher-order system, we can also
factor z-transform polynomials to break down a large system into smaller sections.
Since a cascading system is equivalent to multiplying each individual system transfer
function, the factors of a higher-order polynomial, H(z), would represent component
systems that make up H(z) in a cascade connection. The concept of parallel and
cascade implementation will be further discussed in the realization of IIR filters in
Chapter 6.
Example 4.10: The following LTI system has the transfer function:
H(z) = 1 ÷2z
÷1
÷z
÷3
:
This transfer function can be factored as
H(z) = 1 ÷z
÷1

1 ÷z
÷1
÷z
÷2

= H
1
(z)H
2
(z):
Thus the overall system H(z) can be realized as the cascade of the first-order
system H
1
(z) = 1 ÷z
÷1
and the second-order system H
2
(z) = 1 ÷z
÷1
÷z
÷2
.
4.3.2 Digital Filters
The general I/O difference equation of an FIR filter is given in (3.1.16). Taking the
z-transform of both sides, we have
Y(z) = b
0
X(z) ÷b
1
z
÷1
X(z) ÷ ÷b
L÷1
z
÷(L÷1)
X(z)
= b
0
÷b
1
z
÷1
÷ ÷b
L÷1
z
÷(L÷1)

X(z): (4:3:7)
Therefore the transfer function of the FIR filter is expressed as
H(z) =
Y(z)
X(z)
= b
0
÷b
1
z
÷1
÷ ÷b
L÷1
z
÷(L÷1)
=
¸
L÷1
l=0
b
l
z
÷1
: (4:3:8)
The signal-flow diagram of the FIR filter is shown in Figure 3.6. FIR filters can be
implemented using the I/O difference equation given in (3.1.16), the transfer function
defined in (4.3.8), or the signal-flow diagram illustrated in Figure 3.6.
Similarly, taking the z-transform of both sides of the IIR filter defined in (3.2.18)
yields
Y(z) = b
0
X(z) ÷b
1
z
÷1
X(z) ÷ ÷b
L÷1
z
÷L÷1
X(z) ÷a
1
z
÷1
Y(z) ÷ ÷a
M
z
÷M
Y(z)
=
¸
L÷1
l=0
b
l
z
÷l

X(z) ÷
¸
M
m=1
a
m
z
÷m

Y(z): (4:3:9)
SYSTEMS CONCEPTS 143
By rearranging the terms, we can derive the transfer function of an IIR filter as
H(z) =
Y(z)
X(z)
=
¸
L÷1
l=0
b
l
z
÷l
1 ÷
¸
M
m=1
a
m
z
÷m
=
B(z)
1 ÷A(z)
, (4:3:10)
where B(z) =
¸
L÷1
l=0
b
l
z
÷l
and A(z) =
¸
M
m=1
a
m
z
÷m
. Note that if all a
m
= 0, the IIR filter
given in (4.3.10) is equivalent to the FIR filter described in (4.3.8).
The block diagram of the IIR filter defined in (4.3.10) can be illustrated in Figure
4.5, where A(z) and B(z) are the FIR filters as shown in Figure 3.6. The numerator
coefficients b
l
and the denominator coefficients a
m
are referred to as the feedforward
and feedback coefficients of the IIR filter defined in (4.3.10). A more detailed signal-
flow diagram of an IIR filter is illustrated in Figure 4.6 assuming that M = L ÷1. IIR
filters can be implemented using the I/O difference equation expressed in (3.2.18), the
transfer function given in (4.3.10), or the signal-flow diagram shown in Figure 4.6.
4.3.3 Poles and Zeros
Factoring the numerator and denominator polynomials of H(z), Equation (4.3.10) can
be further expressed as the rational function
x(n) y(n)
y(n−1)
B(z)
A(z)
z
−1
Figure 4.5 IIR filter H(z) consists of two FIR filters A(z) and B(z)
b
0
x(n) y(n)
− a
1
− a
2
− a
M
y(n−M)
y(n−2)
y(n−1)
z
−1
z
−1
b
1
b
2
b
L−1
x(n−L+1)
z
−1
z
−1
Figure 4.6 Detailed signal-flow diagram of IIR filter
144 FREQUENCY ANALYSIS
H(z) =
b
0
a
0
z
M÷L÷1
¸
L÷1
l=1
(z ÷z
l
)
¸
M
m=1
(z ÷p
m
)
, (4:3:11)
where a
0
= 1. Without loss of generality, we let M = L ÷1 in (4.3.11) in order to obtain
H(z) = b
0
¸
M
l=1
(z ÷z
l
)
¸
M
m=1
(z ÷p
m
)
=
b
0
(z ÷z
1
)(z ÷z
2
) (z ÷z
M
)
(z ÷p
1
)(z ÷p
2
) (z ÷p
M
)
: (4:3:12)
The roots of the numerator polynomial are called the zeros of the transfer function H(z).
In other words, the zeros of H(z) are the values of z for which H(z) = 0, i.e., B(z) = 0.
Thus H(z) given in (4.3.12) has M zeros at z = z
1
, z
2
, F F F , z
M
. The roots of the denom-
inator polynomial are called the poles, and there are M poles at z = p
1
, p
2
, F F F , p
M
. The
poles of H(z) are the values of z such that H(z) = ·. The LTI system described in
(4.3.12) is a pole±zero system, while the system described in (4.3.8) is an all-zero system.
The poles and zeros of H(z) may be real or complex, and some poles and zeros may be
identical. When they are complex, they occur in complex-conjugate pairs to ensure that
the coefficients a
m
and b
l
are real.
Example 4.11: Consider the simple moving-average filter given in (3.2.1). Taking
the z-transform of both sides, we have
Y(z) =
1
L
¸
L÷1
l=0
z
÷l
X(z):
Using the geometric series defined in Appendix A.2, the transfer function of the
filter can be expressed as
H(z) =
Y(z)
X(z)
=
1
L
¸
L÷1
l=0
z
÷l
=
1
L
1 ÷z
÷L
1 ÷z
÷1
¸
: (4:3:13)
This equation can be rearranged as
Y(z) = z
÷1
Y(z) ÷
1
L
X(z) ÷z
÷L
X(z)

:
Taking the inverse z-transform of both sides and rearranging terms, we obtain
y(n) = y(n ÷1) ÷
1
L
x(n) ÷x(n ÷L) [ [:
This is an effective way of deriving (3.2.2) from (3.2.1).
SYSTEMS CONCEPTS 145
The roots of the numerator polynomial z
L
÷1 = 0 determine the zeros of H(z)
defined in (4.3.13). Using the complex arithmetic given in Appendix A.3, we have
z
k
= e
j(2p=L)k
, k = 0, 1, F F F , L ÷1: (4:3:14)
Therefore there are L zeros on the unit circle [z[ = 1. Similarly, the poles of H(z) are
determined by the roots of the denominator z
L÷1
(z ÷1). Thus there are L ÷1 poles at
the origin z = 0 and one pole at z = 1. A pole±zero diagram of H(z) given in (4.3.13) for
L = 8 on the complex plane is illustrated in Figure 4.7. The pole±zero diagram provides
an insight into the properties of a given LTI system.
Describing the z-transform H(z) in terms of its poles and zeros will require finding the
roots of the denominator and numerator polynomials. For higher-order polynomials,
finding the roots is a difficult task. To find poles and zeros of a rational function H(z),
we can use the MATLAB function roots on both the numerator and denominator
polynomials. Another useful MATLAB function for analyzing transfer function is
zplane(b, a), which displays the pole±zero diagram of H(z).
Example 4.12: Consider the IIR filter with the transfer function
H(z) =
1
1 ÷z
÷1
÷0:9z
÷2
:
We can plot the pole±zero diagram using the following MATLAB script:
b = [1]; a = [1, ÷1, 0.9];
zplane(b, a);
Similarly, we can plot Figure 4.7 using the following MATLAB script:
b = [1, 0, 0, 0, 0, 0, 0, 0, 1]; a = [1, ÷1];
zplane(b, a);
As shown in Figure 4.7, the system has a single pole at z = 1, which is at the same
location as one of the eight zeros. This pole is canceled by the zero at z = 1. In this case,
the pole±zero cancellation occurs in the system transfer function itself. Since the system
Re[z]
Im[z]
|z|=1
Zero
Pole
Figure 4.7 Pole±zero diagram of the moving-averaging filter, L = 8
146 FREQUENCY ANALYSIS
output Y(z) = X(z)H(z), the pole±zero cancelation may occur in the product of system
transfer function H(z) with the z-transform of the input signal X(z). By proper selection
of the zeros of the system transfer function, it is possible to suppress one or more poles of
the input signal from the output of the system, or vice versa. When the zero is located
very close to the pole but not exactly at the same location to cancel the pole, the system
response has a very small amplitude.
The portion of the output y(n) that is due to the poles of X(z) is called the forced
response of the system. The portion of the output that is due to the poles of H(z) is
called the natural response. If a system has all its poles within the unit circle, then its
natural response dies down as n ÷·, and this is referred to as the transient response. If
the input to such a system is a periodic signal, then the corresponding forced response is
called the steady-state response.
Consider the recursive power estimator given in (3.2.11) as an LTI system H(z) with
input w(n) = x
2
(n) and output y(n) =
”
P
x
(n). As illustrated in Figure 4.8, Equation
(3.2.11) can be rewritten as
y(n) = (1 ÷a)y(n ÷1) ÷aw(n):
Taking the z-transform of both sides, we obtain the transfer function that describes this
efficient power estimator as
H(z) =
Y(z)
W(z)
=
a
1 ÷(1 ÷a)z
÷1
: (4:3:15)
This is a simple first-order IIR filter with a zero at the origin and a pole at z = 1 ÷a. A
pole±zero plot of H(z) given in (4.3.15) is illustrated in Figure 4.9. Note that a = 1=L
results in 1 ÷a = (L ÷1)=L, which is slightly less than 1. When L is large, i.e., a longer
window, the pole is closer to the unit circle.
x(n)
(•)
2
w(n) = x
2
(n)
H(z)
y(n) = P
x
(n)
ˆ
Figure 4.8 Block diagram of recursive power estimator
Re[z]
Im[z]
|z| = 1
Zero
Pole
Figure 4.9 Pole±zero diagram of the recursive power estimator
SYSTEMS CONCEPTS 147
An LTI systemH(z) is stable if and only if all the poles are inside the unit circle. That is,
[p
m
[ < 1 for all m: (4:3:16)
In this case, lim
n÷·
¦h(n)¦ = 0. In other words, an LTI system is stable if and only if the
unit circle is inside the ROC of H(z).
Example 4.13: Given an LTI system with transfer function
H(z) =
z
z ÷a
:
There is a pole at z = a. From Table 4.3, we show that
h(n) = a
n
, n _ 0:
When [a[ > 1, i.e., the pole at z = a is outside the unit circle, we have
lim
n÷·
h(n) ÷·:
that is an unstable system. However, when [a[ < 1, the pole is inside the unit circle,
we have
lim
n÷·
h(n) ÷0,
which is a stable system.
The power estimator described in (4.3.15) is stable since the pole at 1 ÷a
= (L ÷1)=L < 1 is inside the unit circle. A system is unstable if H(z) has pole(s) outside
the unit circle or multiple-order pole(s) on the unit circle. For example, if H(z)
= z=(z ÷1)
2
, then h(n) = n, which is unstable. A system is marginally stable, or oscilla-
tory bounded, if H(z) has first-order pole(s) that lie on the unit circle. For example, if
H(z) = z=(z ÷1), then h(n) = (÷1)
n
, n _ 0.
4.3.4 Frequency Responses
The frequency response of a digital system can be readily obtained from its transfer
function. If we set z = e
j!
in H(z), we have
H(z)

z = e
j!
=
¸
·
n=÷·
h(n)z
÷n

z = e
j!
=
¸
·
n=÷·
h(n)e
÷j!n
= H(!): (4:3:17)
Thus the frequency response of the system is obtained by evaluating the transfer
function on the unit circle [z[ = [e
j!
[ = 1. As summarized in Table 3.1, the digital
frequency ! = 2pf =f
s
is in the range ÷p _ ! _ p.
The characteristics of the system can be described using the frequency response of the
frequency !. In general, H(!) is a complex-valued function. It can be expressed in polar
form as
148 FREQUENCY ANALYSIS
H(!) = [H(!)[e
jf(!)
, (4:3:18)
where [H(!)[ is the magnitude (or amplitude) response and f(!) is the phase shift
(phase response) of the system at frequency !. The magnitude response [H(!)[ is an
even function of !, and the phase response f(!) is an odd function of !. We only need
to know that these two functions are in the frequency region 0 _ ! _ p. The quantity
[H(!)[
2
is referred to as the squared-magnitude response. The value of [H(!
0
)[ for a
given H(!) is called the system gain at frequency !
0
.
Example 4.14: The simple moving-average filter expressed as
y(n) =
1
2
[x(n) ÷x(n ÷1)[, n _ 0
is a first-order FIR filter. Taking the z-transform of both sides and re-arranging
the terms, we obtain
H(z) =
1
2
1 ÷z
÷1

:
From (4.3.17), we have
H(!) =
1
2
1 ÷e
÷j!

=
1
2
(1 ÷cos ! ÷j sin !),
[H(!)[
2
= ¦Re[H(!)[¦
2
÷¦Im[H(!)[¦
2
=
1
2
(1 ÷cos !),
f(!) = tan
÷1
Im[H(!)[
Re[H(!)[

= tan
÷1
÷sin !
1 ÷cos !

:
From Appendix A.1, we have
sin ! = 2 sin
!
2

cos
!
2

and cos ! = 2 cos
2
!
2

÷1:
Therefore the phase response is
f(!) = tan
÷1
÷tan
!
2

= ÷
!
2
:
As discussed earlier, MATLAB is an excellent tool for analyzing signals in the
frequency domain. For a given transfer function, H(z), expressed in a general form in
(4.3.10), the frequency response can be analyzed with the MATLAB function
[H, w]= freqz(b, a, N);
which returns the N-point frequency vector w and the N-point complex frequency
response vector H, given its numerator and denominator coefficients in vectors b and
a, respectively.
SYSTEMS CONCEPTS 149
Example 4.15: Consider the difference equation of IIR filter defined as
y(n) = x(n) ÷y(n ÷1) ÷0:9y(n ÷2): (4:3:19a)
This is equivalent to the IIR filter with the transfer function
H(z) =
1
1 ÷z
÷1
÷0:9z
÷2
: (4:3:19b)
The MATLAB script to analyze the magnitude and phase responses of this IIR
filter is listed (exam 4_15.m in the software package) as follows:
b = [1]; a = [1, ÷1, 0.9];
[H, w]= freqz(b, a, 128);
magH = abs(H); angH = angle(H);
subplot(2, 1, 1), plot(magH), subplot(2, 1, 2), plot(angH);
The MATLAB function abs(H)returns the absolute value of the elements of H
and angle(H)returns the phase angles in radians.
A simple, but useful, method of obtaining the brief frequency response of an LTI
system is based on the geometric evaluation of its pole±zero diagram. For example,
consider a second-order IIR filter expressed as
H(z) =
b
0
÷b
1
z
÷1
÷b
2
z
÷2
1 ÷a
1
z
÷1
÷a
2
z
÷2
=
b
0
z
2
÷b
1
z ÷b
2
z
2
÷a
1
z ÷a
2
: (4:3:20)
The roots of the characteristic equation
z
2
÷a
1
z ÷a
2
= 0 (4:3:21)
are the poles of the filter, which may be either real or complex. For complex poles,
p
1
= re
jy
and p
2
= re
÷jy
, (4:3:22)
where r is radius of the pole and y is the angle of the pole. Therefore Equation (4.3.20)
becomes
z ÷re
jy

z ÷re
÷jy

= z
2
÷2r cos y ÷r
2
= 0: (4:3:23)
Comparing this equation with (4.3.21), we have
r =

a
2

and y = cos
÷1
÷a
1
=2r ( ): (4:3:24)
The filter behaves as a digital resonator for r close to unity. The system with a pair of
complex-conjugated poles as given in (4.3.22) is illustrated in Figure 4.10.
150 FREQUENCY ANALYSIS
r
r
q
q
Im[z]
Re[z]
|z| = 1
Figure 4.10 A second-order IIR filter with complex-conjugated poles
Re[z]
Im[z]
z = e
jw
z = −1
p
1
p
2
V
2
V
1
U
1
U
2
z
1
z
2
Figure 4.11 Geometric evaluation of the magnitude response from the pole±zero diagram
Similarly, we can obtain two zeros, z
1
and z
2
, by evaluating b
0
z
2
÷b
1
z ÷b
2
= 0. Thus
the transfer function defined in (4.3.20) can be expressed as
H(z) =
b
0
z ÷z
1
( ) z ÷z
2
( )
z ÷p
1
( ) z ÷p
2
( )
: (4:3:25)
In this case, the frequency response is given by
H(!) =
b
0
e
j!
÷z
1
( ) e
j!
÷z
2
( )
e
j!
÷p
1
( ) e
j!
÷p
2
( )
: (4:3:26)
Assuming that b
0
= 1, the magnitude response of the system can be shown as
[H(!)[ =
U
1
U
2
V
1
V
2
, (4:3:27)
where U
1
and U
2
represent the distances from the zeros z
1
and z
2
to the point z = e
j!
,
and V
1
and V
2
are the distances of the poles p
1
and p
2
, to the same point as illustrated in
Figure 4.11. The complete magnitude response can be obtained by evaluating [H(!)[ as
the point z moves from z = 0 to z = ÷1 on the unit circle. As the point z moves closer to
the pole p
1
, the length of the vector V
1
decreases, and the magnitude response increases.
SYSTEMS CONCEPTS 151
When the pole p
1
is close to the unit circle, V
1
becomes very small when z is on the same
radial line with pole p
1
(! = y). The magnitude response has a peak at this resonant
frequency. The closer r is to the unity, the sharper the peak. The digital resonator is an
elementary bandpass filter with its passband centered at the resonant frequency y. On
the other hand, as the point z moves closer to the zero z
1
, the zero vector U
1
decreases as
does the magnitude response. The magnitude response exhibits a peak at the pole angle,
whereas the magnitude response falls to the valley at the zero.
4.4 Discrete Fourier Transform
In Section 4.1, we developed the Fourier series representation for continuous-time
periodic signals and the Fourier transform for finite-energy aperiodic signals. In this
section, we will repeat similar developments for discrete-time signals. The discrete-time
signals to be represented in practice are of finite duration. An alternative transformation
called the discrete Fourier transform (DFT) for a finite-length signal, which is discrete
in frequency, also will be introduced in this section.
4.4.1 Discrete-Time Fourier Series and Transform
As discussed in Section 4.1, the Fourier series representation of an analog periodic
signal of period T
0
consists of an infinite number of frequency components, where the
frequency spacing between two successive harmonics is 1=T
0
. However, as discussed in
Chapter 3, the frequency range for discrete-time signals is defined over the interval
(÷p, p). A periodic digital signal of fundamental period N samples consists of frequency
components separated by 2p=N radians, or 1/N cycles. Therefore the Fourier series
representation of the discrete-time signal will contain up to a maximum of N frequency
components.
Similar to (4.1.1), given a periodic signal x(n) with period N such that
x(n) = x(n ÷N), the Fourier series representation of x(n) is expressed as
x(n) =
¸
N÷1
k=0
c
k
e
jk(2p=N)n
, (4:4:1)
which consists of N harmonically related exponentials functions e
jk(2p=N)n
for
k = 0, 1, F F F , N ÷1. The Fourier coefficients, c
k
, are defined as
c
k
=
1
N
¸
N÷1
n=0
x(n)e
÷jk(2p=N)n
: (4:4:2)
These Fourier coefficients form a periodic sequence of fundamental period N such
that
c
k÷iN
= c
k
, (4:4:3)
152 FREQUENCY ANALYSIS
where i is an integer. Thus the spectrum of a periodic signal with period N is a periodic
sequence with the same period N. The single period with frequency index
k = 0, 1, F F F , N ÷1 corresponds to the frequency range 0 _ f _ f
s
or 0 _ F _ 1.
Similar to the case of analog aperiodic signals, the frequency analysis of discrete-time
aperiodic signals involves the Fourier transform of the time-domain signal. In previous
sections, we have used the z-transform to obtain the frequency characteristics of discrete
signals and systems. As shown in (4.3.17), the z-transform becomes the evaluation of the
Fourier transform on the unit circle z = e
j!
. Similar to (4.1.10), the Fourier transform
of a discrete-time signal x(n) is defined as
X(!) =
¸
·
n=÷·
x(n)e
÷j!n
: (4:4:4)
This is called the discrete-time Fourier transform (DTFT) of the discrete-time signal
x(n).
It is clear that X(!) is a complex-valued continuous function of frequency !, and
X(!) is periodic with period 2p. That is,
X(! ÷2pi) = X(!): (4:4:5)
Thus the frequency range for a discrete-time signal is unique over the range (÷p, p) or
(0, 2p). For real-valued x(n), X(!) is complex-conjugate symmetric. That is,
X(÷!) = X
+
(!): (4:4:6)
Similar to (4.1.9), the inverse discrete-time Fourier transform of X(!) is given by
x(n) =
1
2p

p
÷p
X(!)e
j!n
d!: (4:4:7)
Consider an LTI system H(z) with input x(n) and output y(n). From (4.3.1) and
letting z = e
j!
, we can express the output spectrum of system in terms of its frequency
response and the input spectrum. That is,
Y(!) = H(!)X(!), (4:4:8)
where X(!) and Y(!) are the DTFT of the input x(n) and output y(n), respectively.
Similar to (4.3.18), we can express X(!) and Y(!) as
X(!) = [X(!)[e
jf
x
(!)
, (4:4:9a)
Y(!) = [Y(!)[e
jf
y
(!)
: (4:4:9b)
Substituting (4.3.18) and (4.4.9) into (4.4.8), we have
[Y(!)[e
jf
y
(!)
= [H(!)|X(!)[e
j[f(!)÷f
x
(!)[:
(4:4:10)
DISCRETE FOURIER TRANSFORM 153
This equation shows that
[Y(!)[ = [H(!)|X(!)[ (4:4:11)
and
f
y
(!) = f(!) ÷f
x
(!): (4:4:12)
Therefore the output magnitude spectrum [Y(!)[ is the product of the magni-
tude response [H(!)[ and the input magnitude spectrum [X(!)[. The output phase
spectrum f
y
(!) is the sum of the system phase response f(!) and the input
phase spectrum f
x
(!).
For example, if the input signal is the sinusoidal signal at frequency !
0
expressed as
x(n) = Acos(!
0
n), (4:4:13)
its steady-state response can be expressed as
y(n) = A[H(!
0
)[ cos[!
0
÷f(!
0
)[, (4:4:14)
where [H(!
0
)[ is the system amplitude gain at frequency !
0
and f(!
0
) is the phase shift
of the system at frequency !
0
. Therefore it is clear that the sinusoidal steady-state
response has the same frequency as the input, but its amplitude and phase angle are
determined by the system's magnitude response [H(!)[ and phase response f(!) at any
given frequency !
0
.
4.4.2 Aliasing and Folding
As discussed in Section 4.1, let x(t) be an analog signal, and let X( f ) be its Fourier
transform, defined as
X( f ) =

·
÷·
x(t)e
÷j2pft
dt, (4:4:15)
where f is the frequency in Hz. The sampling of x(t) with sampling period T yields the
discrete-time signal x(n). Similar to (4.4.4), the DTFT of x(n) can be expressed as
X(F) =
¸
·
n=÷·
x(n)e
÷j2pFn
: (4:4:16)
The periodic sampling imposes a relationship between the independent variables t and
n in the signals x(t) and x(n) as
t = nT =
n
f
s
: (4:4:17)
154 FREQUENCY ANALYSIS
This relationship in time domain implies a corresponding relationship between the
frequency variable f and F in X( f ) and X(F), respectively. Note that F = f =f
s
is the
normalized digital frequency defined in (3.1.8). It can be shown that
X(F) =
1
T
¸
·
k=÷·
X( f ÷kf
s
): (4:4:18)
This equation states that X(F) is the sum of all repeated values of X( f ), scaled by 1/T,
and then frequency shifted to kf
s
. It also states that X(F) is a periodic function with
period T = 1=f
s
. This periodicity is necessary because the spectrum X(F) of the discrete-
time signal x(n) is periodic with period F = 1 or f = f
s
. Assume that a continuous-time
signal x(t) is bandlimited to f
M
, i.e.,
[X( f )[ = 0 for [ f [ _ f
M
, (4:4:19)
where f
M
is the bandwidth of signal x(t). The spectrum is 0 for [ f [ _ f
M
as shown in
Figure 4.12(a).
X( f )
0 − f
M
− f
s
− f
s
f
s
− f
M
f
M
f
s
f
M
f
(a)
X( f / f
s
)
X( f / f
s
)
0
− f
s
− f
M
f
M
f
s
0
2
f
s
2
2
− f
s
2
f
(b)
f
(c)
Figure 4.12 Spectrum replication caused by sampling: (a) spectrum of analog bandlimited
signal x(t), (b) sampling theorem is satisfied, and (c) overlap of spectral components
DISCRETE FOURIER TRANSFORM 155
The effect of sampling is that it extends the spectrum of X( f ) repeatedly on both sides
of the f-axis, as shown in Figure 4.12. When the sampling rate f
s
is selected to be greater
than 2 f
M
, i.e., if f
M
_ f
s
=2, the spectrum X( f ) is preserved in X(F) as shown in Figure
4.12(b). Therefore when f
s
_ 2f
M
, we have
X(F) =
1
T
X( f ) for [F[ _
1
2
or [ f [ _ f
N
, (4:4:20)
where f
N
= f
s
=2 is called the Nyquist frequency. In this case, there is no aliasing, and the
spectrum of the discrete-time signal is identical (within the scale factor 1/T) to the
spectrum of the analog signal within the fundamental frequency range [ f [ _ f
N
or
[F[ _ 1=2. The analog signal x(t) can be recovered from the discrete-time signal x(n)
by passing it through an ideal lowpass filter with bandwidth f
M
and gain T. This
fundamental result is the sampling theorem defined in (1.2.3). This sampling theorem
states that a bandlimited analog signal x(t) with its highest frequency (bandwidth) being
f
M
can be uniquely recovered from its digital samples, x(n), provided that the sampling
rate f
s
_ 2f
M
.
However, if the sampling rate is selected such that f
s
< 2f
M
, the shifted replicas of
X( f ) will overlap in X(F), as shown in Figure 4.12(c). This phenomenon is called
aliasing, since the frequency components in the overlapped region are corrupted when
the signal is converted back to the analog form. As discussed in Section 1.1, we used an
analog lowpass filter with cut-off frequency less than f
N
before the A/D converter in
order to prevent aliasing. The goal of filtering is to remove signal components that may
corrupt desired signal components below f
N
. Thus the lowpass filter is called the
antialiasing filter.
Consider two sinewaves of frequencies f
1
= 1 Hz and f
2
= 5 Hz that are sampled at
f
s
= 4 Hz, rather than at 10 Hz according to the sampling theorem. The analog wave-
forms are illustrated in Figure 4.13(a), while their digital samples and reconstructed
waveforms are illustrated in Figure 4.13(b). As shown in the figures, we can reconstruct
the original waveform from the digital samples for the sinewave of frequency f
1
= 1 Hz.
However, for the original sinewave of frequency f
2
= 5 Hz, the reconstructed signal
is identical to the sinewave of frequency 1 Hz. Therefore f
1
and f
2
are said to be aliased to
one another, i.e., they cannot be distinguished by their discrete-time samples.
In general, the aliasing frequency f
2
related to f
1
for a given sampling frequency f
s
can
be expressed as
f
2
= if
s
±f
1
, i _ 1: (4:4:21)
For example, if f
1
= 1 Hz and f
s
= 4 Hz, the set of aliased f
2
corresponding to f
1
is given
by
f
2
= i 4 ±1, i = 1, 2, 3, F F F
= 3, 5, 7, 9, F F F (4:4:22)
The folding phenomenon can be illustrated as the aliasing diagram shown in Figure
4.14. From the aliasing diagram, it is apparent that when aliasing occurs, aliasing
frequencies in x(t) that are higher than f
N
will fold over into the region 0 _ f _ f
N
.
156 FREQUENCY ANALYSIS
x(t), f
1
= 1 Hz
t, second
1
x(n)
x(t)
x(t), f
2
= 5 Hz
t
1
x(n)
x(t)
(a)
x(n), f
1
= 1 Hz x(n), f
2
= 5 Hz
n n
x(t)
x(t)
x(n)
x(n)
(b)
Figure 4.13 Example of the aliasing phenomenon: (a) original analog waveforms and digital
samples for f
1
= 1 Hz and f
2
= 5 Hz, and (b) digital samples of f
1
= 1 Hz and f
2
= 5 Hz and
reconstructed waveforms
0
f
1
= 1
f
2
= 5
f
2
= 3 2f
N
= 4
f
N
= 2
3f
N
= 6
f
2
= 7 4f
N
= 8
Figure 4.14 An example of aliasing diagram for f
1
= 1 Hz and f
s
= 4 Hz
4.4.3 Discrete Fourier Transform
To perform frequency analysis of a discrete-time signal x(n), we convert the time-
domain signal into frequency domain using the DTFT defined in (4.4.4). However,
X(!) is a continuous function of frequency, and it also requires an infinite number of
time-domain samples x(n) for calculation. Thus the DTFT is not computationally
feasible using DSP hardware. In this section, we briefly introduce the discrete Fourier
transform (DFT) of the finite-duration sequence x(n) by sampling its spectrum X(!).
We will further discuss DFT in detail in Chapter 7.
DISCRETE FOURIER TRANSFORM 157
The discrete Fourier transform for transforming a time-domain signal
¦x(0), x(1), x(2), F F F , x(N ÷1)¦ into frequency-domain samples ¦X(k)¦ of length N
is expressed as
X(k) =
¸
N÷1
n=0
x(n)e
÷j
2p
N
( )kn
, k = 0, 1, F F F , N ÷1, (4:4:23)
where n is the time index, k is the frequency index, and X(k) is the kth DFT coefficient.
The inverse discrete Fourier transform (IDFT) is defined as
x(n) =
1
N
¸
N÷1
k=0
X(k)e
j
2p
N
( )kn
, n = 0, 1, F F F , N ÷1: (4:4:24)
Equation (4.4.23) is called the analysis equation for calculating the spectrum from the
signal, and (4.4.24) is called the synthesis equation used to reconstruct the signal from its
spectrum. This pair of DFT and IDFT equations holds for any discrete-time signal that
is periodic with period N.
When we define the twiddle factor as
W
N
= e
÷j
2p
N
( )
, (4:4:25)
the DFT defined in (4.4.23) can be expressed as
X(k) =
¸
N÷1
n=0
x(n)W
kn
N
, k = 0, 1, F F F , N ÷1: (4:4:26)
Similarly, the IDFT defined in (4.4.24) can be expressed as
x(n) =
1
N
¸
N÷1
k=0
X(k)W
÷kn
N
, n = 0, 1, F F F , N ÷1: (4:4:27)
Note that W
N
is the Nth root of unity since e
÷j2p
= 1. Because the W
kn
N
are N-periodic, the
DFTcoefficients are N-periodic. The scalar 1/Nthat appears in the IDFTin (4.4.24) does
not appear in the DFT. However, if we had chosen to define the DFTwith the scalar 1/N,
it would not have appeared in the IDFT. Both forms of these definitions are equivalent.
Example 4.16: In this example, we develop a user-written MATALB function to
implement DFT computations. MATLAB supports complex numbers indicated
by the special functions i and j.
Consider the following M-file (dft.m in the software package):
function [Xk]= dft(xn, N)
% Discrete Fourier transform function
% [Xk]= dft(xn, N)
% where xn is the time-domain signal x(n)
158 FREQUENCY ANALYSIS
% N is the length of sequence
% Xk is the frequency-domain X(k)
n = [0:1:N÷1];
k = [0:1:N÷1];
WN = exp(÷j*2*pi/N); % Twiddle factor
nk = n'*k; % N by N matrix
WNnk = WN.^nk; % Twiddle factor matrix
Xk = xn*WNnk; % DFT
In this M-file, the special character ' (prime or apostrophe) denotes the transpose
of a matrix. The exam4_16.m (included in the software package) with the
following statements:
n = [0:127]; N = 128;
xn = 1.5*sin(0.2*pi*n+0.25*pi);
Xk = dft(xn, N);
semilogy(abs(Xk));
axis([0 63 0 120]);
will display the magnitude spectrum of sinewave x(n) in logarithmic scale, and the
x-axis shows only the range from 0 to p.
4.4.4 Fast Fourier Transform
The DFT and IDFT play an important role in many DSP applications including linear
filtering, correlation analysis, and spectrum analysis. To compute one of the X(k)
coefficients in (4.4.23), we need N complex multiplications and N ÷1 complex add-
itions. To generate N coefficients, we need N
2
multiplications and N
2
÷N additions.
The DFT can be manipulated to obtain a very efficient algorithm to compute it.
Efficient algorithms for computing the DFT are called the fast Fourier transform
(FFT) algorithms, which require a number of operations proportional to N log
2
N
rather than N
2
. The development, implementation, and application of FFT will be
further discussed in Chapter 7.
MATLAB provides the built-in function fft(x), or fft(x, N)to compute the DFT
of the signal vector x. If the argument N is omitted, then the length of the DFT is the
length of x. When the sequence length is a power of 2, a high-speed radix-2 FFT
algorithm is employed. The MATLAB function fft(x, N)performs N-point FFT. If
the length of x is less than N, then x is padded with zeros at the end. If the length of x is
greater than N, fft truncates the sequence x and performs DFT of the first N samples
only. MATLAB also provides ifft(x)to compute the IDFT of the vector x, and
ifft(x, N)to calculate the N-point IDFT.
The function fft(x, N)generates N DFT coefficients X(k) for k = 0, 1, F F F N ÷1.
The Nyquist frequency ( f
N
= f
s
=2) corresponds to the frequency index k = N=2. The
frequency resolution of the N-point DFT is
Á =
f
s
N
: (4:4:28)
DISCRETE FOURIER TRANSFORM 159
The frequency f
k
(in Hz) corresponding to the index k can be computed by
f
k
= kÁ =
kf
s
N
, k = 0, 1, F F F , N ÷1: (4:4:29)
Since the magnitude spectrum [X(k)[ is an even function of k, we only need to display
the spectrum for 0 _ k _ N=2 (or 0 _ !
k
_ p).
Example 4.17: By considering the sinewave given in Example 3.1, we can generate
the time-domain signal and show the magnitude spectrum of signal by using the
following MATLAB script (exam4_17.m in the software package):
N = 256;
n = [0:N÷1];
omega = 0.25*pi;
xn = 2*sin(omega*n);
Xk = fft(xn, N); % Perform FFT
absXk = abs(Xk); % Compute magnitude spectrum
plot(absXk(1:(N/2))); % Plot from 0 to p
The phase response can be obtained using the MATLAB function phase =
angle(Xk), which returns the phase angles in radians of the elements of complex
vector Xk.
4.5 Applications
In this section, we will introduce two examples of using frequency analysis techniques
for designing simple notch filters and analyzing room acoustics.
4.5.1 Design of Simple Notch Filters
A notch filter contains one or more deep notches (nulls) in its magnitude response. To
create a null in the frequency response at frequency !
0
, we simply introduce a pair of
complex-conjugate zeros on the unit circle at angle !
0
. That is, at
z = e
±j!
0
: (4:5:1)
Thus the transfer function for an FIR notch filter is
H(z) = (1 ÷e
j!
0
z
÷1
)(1 ÷e
÷j!
0
z
÷1
)
= 1 ÷2 cos(!
0
)z
÷1
÷z
÷2
: (4:5:2)
This is the FIR filter of order 2.
The magnitude response of a notch filter described in (4.5.2) having a null at !
0
= 0:2p
can be obtained using the MATLAB script notch.m given in the software package. The
magnitude response of the designed notch filter is illustrated in Figure 4.15.
160 FREQUENCY ANALYSIS
20
10
0
−10
−20
−30
−40
−50
−60
0.0 0.1
M
a
g
n
i
t
u
d
e

(
d
B
)
0.2
Normalized Frequency
0.3 0.4 0.5
Figure 4.15 Magnitude response of a notch filter with zeros only for !
0
= 0:2p
Obviously, the second-order FIR notch filter has a relatively wide bandwidth, which
means that other frequency components around the null are severely attenuated. To
reduce the bandwidth of the null, we may introduce poles into the system. Suppose that
we place a pair of complex-conjugate poles at
z
p
= re
±jy
0
, (4:5:3)
where r and y
0
are radius and angle of poles, respectively. The transfer function for the
resulting filter is
H(z) =
(1 ÷e
j!
0
z
÷1
)(1 ÷e
÷j!
0
z
÷1
)
(1 ÷re
jy
0
z
÷1
)(1 ÷re
÷jy
0
z
÷1
)
=
1 ÷2 cos(!
0
)z
÷1
÷z
÷2
1 ÷2r cos(y
0
)z
÷1
÷r
2
z
÷2
(4:5:4)
The notch filter expressed in (4.5.4) is the second-order IIR filter.
The magnitude response of the filter defined by (4.5.4) is plotted in Figure 4.16 for
!
0
= y
0
= 0:2p, and r = 0:85. When compared with the magnitude response of the FIR
filter shown in Figure 4.15, we note that the effect of the pole is to reduce the bandwidth
of the notch. The MATLAB script notchl.m in the software package is used to
generate Figure 4.16.
Let y
0
be fixed (y
0
= !
0
= 0:2p) and the value of r changed. The magnitude responses
of the filter are shown in Figure 4.17 for r = 0:75, 0.85, and 0.95. Obviously, the closer
the r value to 1 (poles are closer to the unit circle), the narrower the bandwidth. The
MATLAB script notch2.m used to generate Figure 4.17 is available in the software
package.
APPLICATIONS 161
5
−10
−5
0
−15
−20
−25
−30
−35
−40
0.0 0.1
M
a
g
n
i
t
u
d
e

(
d
B
)
0.2
Normalized frequency
0.3 0.4 0.5
Figure 4.16 Magnitude response of a notch filter with zeros and poles, !
0
= y
0
= 0:2p and
r = 0:85
5
−10
−5
r=0.95
r=0.85
r=0.75
0
−15
−20
−25
−30
0.0 0.1
M
a
g
n
i
t
u
d
e

(
d
B
)
0.2
Normalized frequency
0.3 0.4 0.5
Figure 4.17 Magnitude response of notch filter with both zeros and poles, !
0
= y
0
= 0:2p
and different values of r
4.5.2 Analysis of Room Acoustics
A room transfer function (RTF) expresses the transmission characterstics of a sound
between a source (loudspeaker) and a receiver (microphone) in a room, as illustrated in
Figure 4.18, where H(z) denotes the RTF. The output signal of the receiver can be
expressed in the z-domain as
Y(z) = H(z)X(z):
162 FREQUENCY ANALYSIS
x(n) y(n) H(z)
Source Receiver
Room
Figure 4.18 RTF between a source and a receiver in a room
The RTF includes the characterstics of the direct sound and all reflected sounds
(reverberations) in the room.
An efficient model is required to represent the RTF with a few parameters for
reducing memory and computation requirements. The first method for modeling an
RTF is an all-zero model as defined in (4.3.8), with the coefficients corresponding to the
impulse response of the RTF in the time domain. The all-zero model can be realized
with an FIR filter. When the reverberation time is 500 ms, the FIR filter needs 4000
coefficients (at 8 kHz sampling rate) to represent the RTF. Furthermore, the RTF varies
due to changes in the source and receiver positions.
The pole±zero model defined in (4.3.10) can also be used to model RTF. From a
physical point of view, poles represent resonances, and zeros represent time delays and
anti-resonances. Because the poles can represent a long impulse response caused by
resonances with fewer parameters than the zeros, the pole±zero model seems to match a
physical RTF better than the all-zero model. Because the acoustic poles corresponding
to the resonance properties are invariant, the pole±zero model that has constant poles
and variable zeros is cost effective.
It is also possible to use an all-pole modeling of room responses to reduce the
equalizer length. The all-pole model of RTF can be expressed as
H(z) =
1
1 ÷A(z)
=
1
1 ÷
¸
M
m=1
a
m
z
÷m
: (4:5:5)
Acoustic poles correspond to the resonances of a room and do not change even if the
source and receiver positions change or people move. This technique can be applied to
dereverberation of recorded signals, acoustic echo cancellation, etc. In this section, we
show how the MATLAB functions are used to model and analyze the room acoustics.
To evaluate the room transfer function, impulse responses of a rectangular room
(246 ×143 ×111 cubic inches) were measured using the maximum-length sequence
technique. The original data is sampled at 48 kHz, which is then bandlimited to
100±400 Hz and decimated to 1 kHz. An example of room impulse response is shown
in Figure 4.19, which is generated using the following MATLAB script:
load imp.dat;
plot(imp(1:1000)), title(`Room impulse response');
xlabel(`Time'), ylabel(`Amplitude');
APPLICATIONS 163
1.5
1.0
0.5
0.0
−0.5
−1.0
−1.5
0 200 400
A
m
p
l
i
t
u
d
e
600 800 1000
Time
× 10
−4
Figure 4.19 Example of a measure room impulse response
−150
−140
−130
−120
−110
−100
−90
−80
−70
−60
−50
0 100 200
M
a
g
n
i
t
u
d
e
,

d
B
300 400 500
Frequency, Hz
Figure 4.20 Magnitude response of measured RTF
where the room impulse response samples are stored in the ASCII file imp.dat. Both
the MATLAB script imprtf.m and the data file imp.dat are included in the software
package.
We can easily evaluate the magnitude response of the room transfer function using
the MATLAB script magrtf.m available in the software package. The magnitude
response is shown in Figure 4.20.
MATLAB provides a powerful function a = lpc(x, N) to estimate the coefficients
a
m
of an Mth-order all-pole IIR filter. A user-written MATLAB function all_pole.m
that shows the magnitude responses of the measured and modeled RTF is given in the
164 FREQUENCY ANALYSIS
software package. This MATLAB function can be invoked by using the following
commands:
load imp.dat;
all_pole(imp, 120);
The impulse response of the RTF is modeled by the all-pole model defined in (4.5.5) by
using the MATLAB function a = lpc(imp_x, pole_number), where the pole
number M is selected as 120. In order to evaluate the accuracy of the model, the
MATLAB function freqz(1, a, leng, 1000)is used to compute the frequency
response of the estimated model. The magnitude response of the RTF model is then
compared with the measured magnitude response of RTF from the measured room
impulse response. It is shown that the all-pole model matches the peaks better than the
valleys. Note that the higher the model order M, the better the model match can be
obtained.
A pole±zero model for the RTF can be estimated by using Prony's method as follows:
[b, a] = prony(imp_x, nb, na);
where b and a are vectors containing the estimated numerator and denominator
coefficients, and nb and na are orders of numerator b and denominator a.
4.6 Experiments Using the TMS320C55X
In Section 4.4.3, we introduced the DFT and implemented it in the MATLAB
function dft.m. The C program that implements an N-point DFT defined by Equation
(4.4.26) is listed in Table 4.4. The computation of the DFT involves nested loops,
multiplication of complex data samples, and generation of complex twiddle factors. In
the subsequent experiments, we will write assembly routines to implement the DFT
function.
Assuming we have a complex data sample x(n) = x
r
(n) ÷jx
i
(n) and a complex
twiddle factor W
kn
N
= cos(2pkn=N) ÷j sin(2pkn=N) = W
r
÷jW
i
defined in (4.4.25),
the product of x(n) and W
kn
N
can be expressed as
x(n)W
kn
N
= x
r
(n)W
r
÷x
i
(n)W
i
÷j x
i
(n)W
r
÷x
r
(n)W
i
[ [, (4:6:1)
where the subscripts r and i denote the real and imaginary parts of complex variable.
Equation (4.6.1) can be rewritten as
X[n[ = X
r
[n[ ÷jX
i
[n[ (4:6:2)
for n = 0, 1, F F F , N ÷1, where
X
r
[n[ = x
r
(n)W
r
÷x
i
(n)W
i
, (4:6:3a)
X
i
[n[ = x
i
(n)W
r
÷x
r
(n)W
i
: (4:6:3b)
EXPERIMENTS USING THE TMS320C55X 165
The C program listed in Table 4.4 uses two arrays, Xin[2*N] and Xout[2*N], to
represent the complex (real and imaginary) input and output data samples. The
input samples for the experiment are generated using the MATLAB script listed in
Table 4.5.
Table 4.4 List of dft()in C
#define PI 3.1415926536
void dft(float Xin [], float Xout [])
{
int n, k, j;
float angle;
float Xr [N], Xi [N];
float Wn [2];
for(k = 0; k < N; k÷÷)
{
Xr [k]= 0;
Xi [k]= 0;
for(j = 0, n = 0; n < N; n÷÷)
{
angle =(2.0*PI*k*n)/ N;
W [0]= cos(angle);
W [1]= sin(angle);
Xr [k]= Xr [k]÷ Xin [j]*W [0]÷ Xin [j÷1]*W [1];
Xi [k]= Xi [k]÷ Xin [j ÷ 1]*W [0]÷ Xin [j]*W [1];
j ÷= 2;
}
Xout [n÷÷] = Xr [k];
Xout [n÷÷] = Xi [k];
}
}
Table 4.5 Generating signal using MATLAB
fs = 8000; % Sampling frequency in Hz
f1 = 500; % 1st sinewave frequency in Hz
f2 = 1000; % 2nd sinewave frequency in Hz
f3 = 2000; % 3rd sinewave frequency in Hz
n = [0:127]; % n = 0, 1, ..., 127
w1 = 2*pi*f1/fs;
w2 = 2*pi*f2/fs;
w3 = 2*pi*f3/fs;
x = 0.3*sin(w1*n) ÷ 0.3*sin(w2*n) ÷ 0.3*sin(w3*n);
166 FREQUENCY ANALYSIS
Table 4.5 (continued )
fid = fopen(`input.dat', `w'); % Open file input.dat for write
fprintf(fid, `int input [128]= {\n');
fprintf(fid, `%5d, \n', round(x(1:127)*32767));
fprintf(fid, `%5d }; \n', round(x(128)*32767));
fclose(fid); % Close file input.dat
fid = fopen(`input.inc', `w'); % Open file input.inc for write
fprintf(fid, ` .word %5d\n', round(x*32767));
fclose(fid); % Close file input.inc
This program generates 128 data samples. The data is then represented using the
Q15 format for the experiments. They are stored in the data files input.dat and
input.inc, and are included in the software package. The data file input.dat is
used by the C program exp4a.c for Experiment 4A, and the data file input.inc
is used by the assembly routine exp4b.asm for Experiment 4B.
4.6.1 Experiment 4A ± Twiddle Factor Generation
The sine±cosine generator we implemented for the experiments in Chapter 3 can be used
to generate the twiddle factors for comparing the DFT. Recall the assembly function
sine_cos.asm developed in Section 3.8.5. This assembly routine is written as a C-
callable function that follows the C55x C-calling convention. There are two arguments
passed by the function as sine_cos(angle, Wn). The first argument is passed
through the C55x temporary register T0 containing the input angle in radians. The
second argument is passed by the auxiliary register AR0 as a pointer Wn to the memory
locations, where the results of sin(angle) and cos(angle) will be stored upon return.
The following C example shows how to use the assembly sine±cosine generator inside
nested loops to generate the twiddle factors:
#define N 128
#define TWOPIN 0x7FFF6 /* 2pkn/N, N = 128 */
int n, k, angle;
int Wn [2]; /* Wn [0]= cos(angle), Wn [1]= sin(angle) */
for(k = 0; k < N; k÷÷)
{
for(n = 0; n < N; n÷÷)
{
angle = TWOPIN*k*n;
sine_cos(angle, Wn);
}
}
EXPERIMENTS USING THE TMS320C55X 167
The assembly code that calls the subroutine sine_cos is listed as follows:
mov *(angle), T0 ; Pass the first argument in T0
amov #Wn,XAR0 ; Pass the second argument in AR0
call sine_cos ; Call sine_cos subroutine
In Chapter 2, we introduced how to write nested loops using block-repeat and single
repeat instructions. Since the inner loop of the C code dft()contains multiple instruc-
tions, we will use the block-repeat instruction (rptb) to implement both of the inner
and outer loops. The C55x has two block-repeat counters, registers BRC0 and BRC1.
When implementing nested loops, the repeat counter BRC1 must be used as the inner-
loop counter, while the BRC0 should be used as the outer-loop counter. Such an
arrangement allows the C55x to automatically reload the inner-loop repeat counter
BRC1 every time the outer-loop counter being updated. The following is an example of
using BRC0 and BRC1 for nested loops N times:
mov #N÷1, BRC0
mov #N÷1, BRC1
rptb outer_loop-1
(more outer loop instructions...)
rptb inner_loop-1
(more inner loop instructions...)
inner_loop
(more outer loop instructions...)
outer_loop
The calculation of the angle depends on two variables, k and n, as
angle = (2p=N)kn: (4:6:4)
As defined in Figure 3.23 of Section 3.8.5, the fixed-point representation of value p for
sine±cosine generator is 0x7FFF. The angle used to generate the twiddle factors for
DFT of N = 128 can be expressed as
angle = (2
+
0x7FFF=128)
+
k
+
n = (0x1FF)
+
k
+
n, (4:6:5)
where the inner-loop index n = 0, 1, F F F , N ÷1 and the outer-loop index
k = 0, 1, F F F , N ÷1. The following is the assembly routine that calculates the angles
for N = 128:
N .set 128
TWOPIN .set 0x7FFF 6 ; 2*PI/N, N = 128
.bss Wn, 2 ; Wn [0]= Wr, Wn [1]= Wi
.bss angle, 1 ; Angle for sine÷cosine function
mov #N÷1, BRC0 ; Repeat counter for outer-loop
mov #N÷1, BRC1 ; Repeat counter for inner-loop
mov #0, T2 ; k = T2 = 0
168 FREQUENCY ANALYSIS
rptb outer_loop-1 ; for(k = 0; k < N;k÷÷) {
mov #TWOPIN <#16, AC0 ; hi(AC0)= 2*PI/N
mpy T2, AC0
mov #0, *(angle)
mov AC0, T3 ; angle = 2*PI*k/N
rptb inner_loop-1 ; for(n = 0 ; n < N;n÷÷) {
mov *(angle), T0 ; T0 = 2*PI*k*n/N
mov *(angle), AC0
add T3, AC0
mov AC0, *(angle) ; Update angle
amov #Wn, XAR0 ; AR0 is the pointer to Wn
call sine_cos ; sine_cos(angle, Wn)
inner_loop
add #1,T2
outer_loop
4.6.2 Experiment 4B ± Complex Data Operation
For the experiment, the complex data and twiddle factor vectors are arranged
in order of the real and the imaginary pairs. That is, the input array
Xin [2N]= ¦X
r
, X
i
, X
r
, X
i
, F F F¦ and the twiddle factor [2]= ¦W
r
, W
i
¦. The computa-
tion of (4.6.3) is implemented in C as follows:
Xr [n]= 0;
Xi [n]= 0;
for(n = 0; n < N; n÷÷)
{
Xr[n]= Xr[n]÷ Xin[n]*Wr ÷ Xin [n ÷ 1]*Wi;
Xi[n]= Xi[n]÷ Xin[n ÷ 1]*Wr ÷ Xin [n]*Wi;
}
The C55x assembly program implementation of (4.6.3) is listed as follows:
mov #0, AC2
mov #0, AC3
rptb inner_loop-1
macm40 *AR5÷, *AR0, AC2 ; Xin[n]*Wr
macm40 *AR5÷, *AR0÷, AC3 ; Xin[n÷1]*Wr
masm40 *AR5÷, *AR0, AC3 ; Xi[k]= Xin [n÷1]*Wr ÷ Xin [n]*Wi
macm40 *AR5÷, *AR0÷, AC2 ; Xr[k]= Xin [n]*Wr ÷ Xin [n÷1]*Wi
inner_loop
Because the DFT function accumulates N intermediate results, the possible overflow
during computation should be considered. The instruction masm40 enables the use of
accumulator guard bits that allow the intermediate multiply±accumulate result to be
handled in a 40-bit accumulator. Finally, we can put all the pieces together to complete
the routine, as listed in Table 4.6.
EXPERIMENTS USING THE TMS320C55X 169
Table 4.6 List of DFT assembly routine dft_128.asm
; DFT_128 ÷ 128-point DFT routine
;
; Entry T0: AR0: pointer to complex input buffer
; AR1: pointer to complex output buffer
; Return: None
;
.def dft_128
.ref sine_cos
N .set 128
TWOPIN .set 0x7fff 6 ; 2*PI/N, N = 128
.bss Wn, 2 ; Wn [0]= Wr, Wn [1]= Wi
.bss angle, 1 ; Angle for sine÷cosine function
.text
_dft_128
pshboth XAR5 ; Save AR5
bset SATD
mov #N÷1, BRC0 ; Repeat counter for outer loop
mov #N÷1, BRC1 ; Repeat counter for inner loop
mov XAR0, XAR5 ; AR5 pointer to sample buffer
mov XAR0, XAR3
mov #0, T2 ; k =T2 = 0
rptb outer_loop-1 ; for(k = 0; k < N; k÷÷) {
mov XAR3, XAR5 ; Reset x[]pointer
mov #TWOPIN <#16, AC0 ; hi(AC0) = 2*PI/N
mpy T2, AC0
mov #0, AC2 ; Xr[k]= 0
mov #0, AC3 ; Xi[k]= 0
mov #0, *(angle)
mov AC0, T3 ; angle = 2*PI*k/N
rptb inner_loop-1 ; for(n = 0; n < N; n÷÷) {
mov *(angle), T0 ; T0 = 2*PI*k*n/N
mov *(angle), AC0
add T3, AC0
mov AC0, *(angle) ; Update angle
amov #Wn, XAR0 ; AR0 is the pointer to Wn
call _sine_cos ; sine_cos(angle, Wn)
bset SATD ; sine_cos turn off FRCT & SATD
macm40 *AR5÷, *AR0, AC2 ; XR [k]÷ Xin [n]*Wr
macm40 *AR5÷, *AR0÷, AC3 ; XI [k]÷ Xin [n ÷ 1]*Wr
masm40 *AR5÷, *AR0, AC3 ; XI[k]÷ Xin[n ÷ 1]*Wr ÷ Xin[n]*Wi
macm40 *AR5÷, *AR0÷, AC2 ; XR[k]÷ Xin[n]*Wr ÷ Xin[n ÷ 1]*Wi
inner_loop ; } end of inner loop
mov hi(AC2 <#÷5), *AR1÷
mov hi(AC3 <#÷5), *AR1÷
add #1, T2
outer_loop ; } end of outer loop
170 FREQUENCY ANALYSIS
Table 4.6 (continued )
popboth XAR5
bclr SATD
ret
.end
Table 4.7 List of C program exp4a.c
/* Experiment 4A ÷ exp4a.c */
#include "input.dat"
#define N 128
extern void dft_128(int *, int *);
extern void mag_128(int *, int *);
int Xin [2*N];
int Xout [2*N];
int Spectrum [N];
void main()
{
int i, j;
for(j = 0, i = 0; i < N; i÷÷)
{
Xin [j÷÷]= input [i]; /* Get real sample */
Xin [j÷÷]= 0; /* Imaginary sample = 0 */
}
dft_128(Xin, Xout); /* DFT routine */
mag_128(Xout, Spectrum); /* Compute spectrum */
}
4.6.3 Experiment 4C ± Implementation of DFT
We will complete the DFT routine of N = 128 and test it in this section. The C program
listed in Table 4.7 calls for the assembly routine dft_128()to compute the 128-point
DFT.
The data file, input.dat, is an ASCII file that contains 128 points of data sampled
at 8 kHz, and is available in the software package. First, the program composes the
complex input data array Xin [2*N]by zero-filling the imaginary parts. Then the DFT
is carried out by the subroutine dft_128(). The 128 complex DFT samples are stored
in the output data array Xout [2*N]. The subroutine mag_128()at the end of the
program is used to compute the squared magnitude spectrum of the 128 complex DFT
samples from the array Xout [2*N]. The magnitude is then stored in the array called
Spectrum[N], which will be used for graphic display later. The assembly routine,
mag_128.asm, is listed in Table 4.8.
EXPERIMENTS USING THE TMS320C55X 171
Table 4.8 The list of assembly program mag_128.asm
; Compute the magnitude response of the input
;
; Entry: AR0: input buffer pointer
; AR1: output buffer pointer
; Exit: None
;
.def _mag_128
N .set 128
_mag_128
bset SATD
pshboth XAR5
mov #N÷1, BRC0 ; Set BRC0 for loop N times
mov XAR0, XAR5
bset FRCT
rptblocal mag_loop-1
mpym *AR0÷, *AR5÷, AC0 ; Xr [i]*Xr [i]
macm *AR0÷, *AR5÷, AC0 ; Xr [i]*Xr [i]÷ Xi[i]*Xi [i]
mov hi(saturate(AC0)), *AR1÷
mag_loop
popboth XAR5
bclr SATD
bclr FRCT
ret
.end
Perform the following steps for Experiment 4C:
1. Write the C program exp4a.c based on the example (or copy from the software
package) that will complete the following tasks:
(a) Compose the complex input sample to Xin [].
(b) Call the subroutine dft_128()to perform DFT.
(c) Call the subroutine mag_128()to compute the squared magnitude spectrum of
the DFT.
2. Write the assembly routine dft_128.asm for the DFT, and write the assembly
routine mag_128.asm for computing the magnitude spectrum (or copy these files
from the software package).
3. Test and debug the programs. Plot the magnitude spectrum (Spectrum [N]) and
the input samples as shown in Figure 4.21.
4. Profile the DFT routine and record the program and data memory usage. Also,
record the clock cycles used for each subroutine.
172 FREQUENCY ANALYSIS
Figure 4.21 The plots of time-domain input signal (top), the input spectrum (middle) which
shows three peaks are located at frequencies 0.5 kHz, 1 kHz, 2 kHz, and the DFT result (bottom)
4.6.4 Experiment 4D ± Experiment Using Assembly Routines
In the previous experiments, we have written programs using either C programs or C
combined with assembly programs. In some applications, it is desirable to develop the
software using only assembly language. In this section, we will learn how to write and
use assembly routines for an entire experiment.
For previous experiments, we linked the run-time support library rts55.lib to our
programs. This library contains a routine called boot.asm for setting up the processor
from the reset state. The settings include:
1. Initialize extended stack pointer XSP and extended system stack pointer XSSP.
2. Turn the sign extension mode on for arithmetic operations.
3. Turn the 40-bit accumulator mode off, and set the default as 32-bit mode.
4. Turn DU and AU saturate mode off.
5. Turn off the fractional mode.
6. Turn off the circular addressing mode.
When we replace the C function main()with an assembly program, we do not need
to use the run-time support library rts55.lib. However, we have to create an
assembly routine called vectors.asm that will set up the program starting address
and begin the program execution from that point. The following is the list of assembly
code:
EXPERIMENTS USING THE TMS320C55X 173
; vectors.asm
.def rsv
.ref start
.sect "vectors"
rsv .ivec start
where the assembly directive .ivec defines the starting address of the C55x interrupt
vector table.
Interrupts are hardware- or software-driven signals that cause the C55x to suspend
its current program and execute an interrupt service routine (ISR). Once the interrupt
is acknowledged, the C55x executes the branch instruction at the corresponding
interrupt vector table to perform an ISR. There are 32 interrupts and each interrupt
uses 64 bits (4-word) in the C55x vector table. The first 32 bits contain the 24-bit
program address of ISR. The second 32-bit can be ISR instructions. This 32-bit code
will be executed before branching to the ISR. The label start is the entry point of
our experiment program. At power up (or reset), the C55x program counter will be
pointing to the first interrupt vector table, which is a branch instruction to the
label start to begin executing the program. Since the vectors are fixed for the C55x,
we need to map the address of interrupt-vector table .ivec to the program memory at
address 0xFFFF00. The linker command file can be used to map the address of the
vector table.
Because we do not use boot.asm, we are responsible for setting up the system before
we can begin to perform our experiment. The stack pointer must be correctly set before
any subroutine calls (or branch to ISR) can be made. Some of the C55x operation states/
modes should also be set accordingly. The following example shows some of the settings
at the beginning of our program:
stk_size .set 0x100
stack .usect ".stack",stk_size
sysstack .usect ".stack",stk_size
.def start
.sect .text
start
bset SATD
bset SATA
bset SXMD
bclr C54CM
bclr CPL
amov #(stack÷stk_size), XSP
mov #(sysstack÷stk_size), SSP
The label start in the code defines the entry point of the program. Since it will also
be used by the vectors.asm, it needs to be defined as a global label. The first three
bit-set instructions (bset) set up the saturation mode for both the DU and AU and
the sign extension mode. The next two bit-clear instructions (bclr) turn off the
C54x compatibility mode and the C compiler mode. The last two move instructions
(amov/mov) initialize the stack pointers. In this example, the stack size is defined as
0x100 long and starts in the section named .stack in the data memory. When
subroutine calls occur, the 24-bit program counter PC(23:0) will split into two portions.
174 FREQUENCY ANALYSIS
The stack pointer SP is used for storing the lower 16-bit address of the program counter
PC(15:0), and the system stack pointer SSP is used for the upper 8-bit of the
PC(23:16).
In this experiment, we wrote an assembly program exp4b.asm listed in Table 4.9 to
replace the main()function in the C program exp4a.c.
Table 4.9 List of assembly program exp4b.asm
; exp4b.asm: Call dft_128 to compute DFT
;
N .set 128
stk_size .set 0x100
stack .usect ".stack", stk_size
sysstack .usect ".stack", stk_size
_Xin .usect "in_data", 2*N
_Xout .usect "out_data", 2*N
_Spectrum .usect "out_data", N
.sect .data
input .copy input.dat
.def start
.def _Xin, _Xout, _Spectrum
.ref _dft_128, _mag_128
.sect .text
start
bset SATD ; Set up saturation for DU
bset SATA ; Set up saturation for AU
bset SXMD ; Set up sign extension mode
bclr C54CM ; Disable C54x compatibility mode
bclr CPL ; Turn off C compiler mode
amov #(stack÷stk_size), XSP ; Setup DSP stack
mov #(sysstack÷stk_size), SSP ; Setup system stack
mov #N÷1, BRC0 ; Init counter for loop N times
amov #input, XAR0 ; Input data array pointer
amov #Xin, XAR1 ; Xin array pointer
rptblocal complex_data-1 ; Form complex data
mov *AR0÷, *AR1÷
mov #0, *AR1÷
complex_data
amov #_Xin, XAR0 ; Xin array pointer
amov #_Xout, XAR1 ; Xout array pointer
call _dft_128 ; Perform 128-ponts DFT
amov #_Xout, XAR0 ; Xout pointer
amov #_Spectrum, XAR1 ; Spectrum array pointer
call _mag_128 ; Computer squared-mag response
here b here
.end
EXPERIMENTS USING THE TMS320C55X 175
Perform the following steps for Experiment 4D:
1. Write an assembly routine exp4b.asm to replace the C program exp4a.c (or
copy it from the software package).
2. Use the .usect "indata" directive for array Xin[256], .usect "outdata" for
Xout [256]and Spectrum [128], and use sect.code for the program section of
the assembly routine exp4b.asm. Create a linker command file exp4b.cmd and
add the above sections. The code section.code in the program memory starts at
address 0x20400 with a length of 4096 bytes. The indata section starts in the data
memory of word address 0x8000, and its length is 256 words. The outdata section
starts in the data memory with staring address of 0x08800 and has the length of 512
words.
3. Test and debug the programs, verify the memory locations for sections .code,
indata, and outdata, and compare the DFT results with experiment results
obtained in Section 4.6.3.
References
[1] N. Ahmed and T. Natarajan, Discrete-Time Signals and Systems, Englewood Cliffs, NJ: Prentice-
Hall, 1983.
[2] MATLAB User's Guide, Math Works, 1992.
[3] MATLAB Reference Guide, Math Works, 1992.
[4] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, Englewood Cliffs, NJ:
Prentice-Hall, 1989.
[5] S. J. Orfanidis, Introduction to Signal Processing, Englewood Cliffs, NJ: Prentice-Hall, 1996.
[6] J. G. Proakis and D. G. Manolakis, Digital Signal Processing ± Principles, Algorithms, and Applica-
tions, 3rd Ed., Englewood Cliffs, NJ: Prentice-Hall, 1996.
[7] A Bateman and W. Yates, Digital Signal Processing Design, New York: Computer Science Press,
1989.
[8] S. M. Kuo and D. R. Morgan, Active Noise Control Systems ± Algorithms and DSP Implementa-
tions, New York: Wiley, 1996.
[9] H. P. Hsu, Signals and Systems, New York: McGraw-Hill, 1995.
Exercises
Part A
1. Similar to Example 4.1, assume the square wave is expressed as
x(t) =
A, kT
0
_ t _ (k ÷0:5)T
0
0, otherwise,

where k is an integer. Compute the Fourier series coefficients.
176 FREQUENCY ANALYSIS
2. Similar to Example 4.2, compute the Fourier series coefficients for the signal
x(t) = cos(V
0
t):
3. Find and sketch the Fourier transform of the following signals:
(a) The rectangular signal defined as
x(t) =
A, [t[ _ t
0, t > t.

(b) The periodic impulse train defined as
d(t) =
¸
·
k=÷·
d(t ÷kT
0
):
(c) x(t) = d(t).
(d) x(t) = 1.
4. Find the z-transform and ROC of the following sequences:
(a) x(n) = 1, n _ 0.
(b) x(n) = e
÷an
, n _ 0:
(c) x(n) =
÷a
n
, n = ÷1, ÷2, F F F , ÷·
0, n _ 0.

(d) x(n) = sin(!n), n _ 0:
5. The z-transform of an N-periodic sequence can be expressed in terms of the z-transform of its
first period. That is, if
x
1
(n) =
x(n), 0 _ n _ N ÷1
0, elsewhere

denotes the sequence of the first period of the periodic sequence x(n). Show that
X(z) =
X
1
(z)
1 ÷z
÷N
, [z
N
[ > 1,
where X
1
(z) = ZT[x
1
(n)[.
6. Given the periodic signal
x(n) = ¦1, 1, 1, ÷1, ÷1, ÷1, 1, 1, 1, ÷1, F F F¦
of period equal to 6. Find X(z).
EXERCISES 177
7. For a finite length (N) signal, (4.2.2) can be further simplified to
X(z) =
¸
N÷1
n=0
x(n)z
÷n
:
Consider the sequence
x(n) =
a
n
, 0 _ n _ N ÷1
0, otherwise

where a > 0. Find X(z) and plot the poles and zeros of X(z).
8. Find the z-transform and ROC of
x(n) =
1, n < 0
(0:5)
n
, n _ 0.

9. Using partial-fraction-expansion method to find the inverse z-transform of
(a) X(z) =
z
2z
2
÷3z ÷1
, [z[ <
1
2
:
(b) X(z) =
z
3z
2
÷4z ÷1
:
(c) X(z) =
2z
2
(z ÷1)(z ÷2)
2
:
10. Using residue method to find the inverse z-transform of the following functions:
(a) X(z) =
z
2
÷z
(z ÷1)
2
:
(b) X(z) =
z
2
÷2z
(z ÷0:6)
3
:
(c) X(z) =
1
(z ÷0:4)(z ÷1)
2
:
11. The first-order trapezoidal integration rule in numerical analysis is described by the I/O
difference equation
y(n) =
1
2
[x(n) ÷x(n ÷1)[ ÷y(n ÷1), n _ 0:
Treat this rule as an LTI system. Find
(a) The transfer function H(z).
(b) The impulse response h(n).
178 FREQUENCY ANALYSIS
12. Consider the second-order IIR filter
y(n) = a
1
y(n ÷1) ÷a
2
y(n ÷2) ÷x(n), n _ 0:
(a) Compute H(z).
(b) Discuss stability conditions related to coefficients a
1
and a
2
.
13. Determine the stability of the following IIR filters:
(a) H(z) =
z(z ÷1)
(z
2
÷z ÷1)(z ÷0:8)
:
(b) 3y(n) = 3:7y(n ÷1) ÷0:7y(n ÷2) ÷x(n ÷1), n _ 0:
14. In Figure 4.4(b), let H
1
(z) and H
2
(z) are the transfer functions of the two first-order IIR
filters defined as
y
1
(n) = x(n) ÷0:5y
1
(n ÷1), n _ 0
y
2
(n) = x(n) ÷y
2
(n ÷1), n _ 0:
(a) Find the overall transfer function H(z) = H
1
(z) ÷H
2
(z).
(b) Find the output y(n) if the input x(n) = (÷1)
n
, n _ 0.
15. Consider the first-order IIR system
y(n) =
1 ÷a
2
[x(n) ÷x(n ÷1)[ ÷ay(n ÷1), n _ 0,
Find the squared-magnitude response [H(!)[
2
.
16. Consider a moving average filter defined in (3.2.1). Find the magnitude response [H(!)[ and
the phase response f(!).
17. Consider the FIR filter
y(n) = x(n) ÷2x(n ÷1) ÷4x(n ÷2) ÷2x(n ÷3) ÷x(n ÷4), n _ 0:
Find the transfer function H(z) and magnitude response [H(!)[.
18. Derive Equation (4.4.18).
Part B
19. Compute c
k
given in (4.1.4) for A = 1, T
0
= 0:1, and t = 0:05, 0:01, 0:001, and 0.0001.
Using MATLAB function stem to plot c
k
for k = 0, ±1, F F F ±20.
20. Repeat the Problem 19 for A = 1, t = 0:001 and T
0
= 0:005, 0.001, and 0.01.
EXERCISES 179
Part C
21. The assembly program, dft_128.asm, can be further optimized. Use parallel instructions
to improve the DFT performance. Profile the optimized code, and compare the cycle counts
against the profile data obtained in experiment given in Section 4.6.3.
22. Find the clock rate of the TMS320C55x device, and use the profile data to calculate the total
time the DFT routine spent to compute 128-point samples. Can this DFT routine be used for
a real-time application? Why?
23. Why does the C55x DSP have two stack pointers, XSP and XSSP, and how are these
pointers initialized? The C55x also uses RETA and CFCT during subroutine calls. What
are these registers? Create and use a simple assembly program example to describe how the
RETA, CFCT, XSP, and XSSP react to a nested subroutine call (hint: use references from
the CCS help menu).
24. Modify experiments given in Section 4.6.4 to understand how the linker works:
(a) Move all the programs in .text section (exp4b. asm, dft_128.asm and
mag_128.asm) to a new section named .sect "dft_code", which starts at the
program memory of address 0x020400. Adjust the section length if necessary.
(b) Put all the data variables (exp4b. asm, dft_128.asm, and mag_128.asm) under the
section named. usect "dft_vars", which starts at the data memory of address
0x08000. Again, adjust the section length if necessary.
(c) Build and run the project. Examine memory segments in the map file. How does the
linker handle each program and data section?
180 FREQUENCY ANALYSIS
5
Design and Implementation
of FIR Filters
A filter is a system that is designed to alter the spectral content of input signals in
a specified manner. Common filtering objectives include improving signal quality,
extracting information from signals, or separating signal components that have been
previously combined. A digital filter is a mathematical algorithm implemented in hard-
ware, firmware, and/or software that operates on a digital input signal to produce a
digital output signal for achieving filtering objectives. A digital filter can be classified
as being linear or nonlinear, time invariant or varying. This chapter is focused on the
design and implementation of linear, time-invariant (LTI) finite impulse response
(FIR) filters. The time-invariant infinite impulse response (IIR) filters will be discussed
in Chapter 6, and the time-varying adaptive filters are introduced in Chapter 8.
The process of deriving the digital filter transfer function H(z) that satisfies the
given set of specifications is called digital filter design. Although many applications
require only simple filters, the design of more complicated filters requires the use of
sophisticated techniques. A number of computer-aided design tools (such as MATLAB)
are available for designing digital filters. Even though such tools are widely available,
we should understand the basic characteristics of digital filters and familiar with
techniques used for implementing digital filters. Many DSP books devote substantial
efforts to the theory of designing digital filters, especially approximation methods,
reflecting the considerable work that has been done for calculating and optimizing filter
coefficients.
5.1 Introduction to Digital Filters
As discussed in previous chapters, filters can be divided into two categories: analog
filters and digital filters. Similar specifications are used for both analog and digital
filters. In this chapter, we will discuss digital filters exclusively. The digital filters
are assumed to have a single input x(n), and a single output y(n). Analog filters are
used as design prototypes for digital IIR filters, and will be briefly introduced in
Chapter 6.
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
5.1.1 Filter Characteristics
A digital filter is said to be linear if the output due to the application of input,
x(n) = a
1
x
1
(n) ÷a
2
x
2
(n), (5:1:1)
is equal to
y(n) = a
1
y
1
(n) ÷a
2
y
2
(n), (5:1:2)
where a
1
and a
2
are arbitrary constants, and y
1
(n) and y
2
(n) are the filter outputs due to
the application of the inputs x
1
(n) and x
2
(n), respectively. The important property of
linearity is that in the computation of y(n) due to x(n), we may decompose x(n) into a
summation of simpler components x
i
(n). We then compute the response y
i
(n) due to
input x
i
(n). The summation of y
i
(n) will be equal to the output y(n). This property is
also called the superposition.
A time-invariant system is a system that remains unchanged over time. A digital filter
is time-invariant if the output due to the application of delayed input x(n ÷m) is equal
to the delayed output y(n ÷m), where m is a positive integer. It means that if the input
signal is the same, the output signal will always be the same no matter what instant the
input signal is applied. It also implies that the characteristics of a time-invariant filter
will not change over time.
A digital filter is causal if the output of the filter at time n
0
does not depend on the
input applied after n
0
. It depends only on the input applied at and before n
0
. On the
contrary, the output of a non-causal filter depends not only on the past input, but also
on the future input. This implies that a non-causal filter is able to predict the input that
will be applied in the future. This is impossible for any real physical filter.
Linear, time-invariant filters are characterized by magnitude response, phase
response, stability, rise time, settling time, and overshoot. Magnitude response specifies
the gains (amplify, pass, or attenuate) of the filter at certain frequencies, while phase
response indicates the amount of phase changed by the filter at different frequencies.
Magnitude and phase responses determine the steady-state response of the filter. For an
instantaneous change in input, the rise time specifies an output-changing rate. The
settling time describes an amount of time for the output to settle down to a stable
value, and the overshoot shows if the output exceeds the desired output value. The rise
time, the settling time, and the overshoot specify the transient response of the filter in
the time domain.
A digital filter is stable if, for every bounded input signal, the filter output is bounded.
A signal x(n) is bounded if its magnitude [x(n)[ does not go to infinity. A digital filter
with the impulse response h(n) is stable if and only if
¸
·
n=0
[h(n)[ < ·: (5:1:3)
Since an FIR filter has only a finite number of non-zero h(n), the FIR filter is always
stable. Stability is critical in DSP implementations because it guarantees that the filter
182 DESIGN AND IMPLEMENTATION OF FIR FILTERS
output will never grow beyond bounds, thus avoiding numerical overflow in computing
the convolution sums.
As mentioned earlier, filtering is a process that passes certain frequency components
in a signal through the system and attenuates other frequency components. The range of
frequencies that is allowed to pass through the filter is called the passband, and the
range of frequencies that is attenuated by the filter is called the stopband. If a filter is
defined in terms of its magnitude response, there are four different types of filters:
lowpass, highpass, bandpass, and bandstop filters. Each ideal filter is characterized by a
passband over which frequencies are passed unchanged (except with a delay) and a
stopband over which frequencies are rejected completely. The two-level shape of the
magnitude response gives these filters the name brickwall. Ideal filters help in analyzing
and visualizing the processing of actual filters employed in signal processing. Achieving
an ideal brickwall characteristic is not feasible, but ideal filters are useful for concep-
tualizing the impact of filters on signals.
As discussed in Chapter 3, there are two basic types of digital filters: FIR filters and
IIR filters. An FIR filter of length L can be represented with its impulse response h(n)
that has only L non-zero samples. That is, h(n) = 0 for all n _ L. An FIR filter is also
called a transversal filter. Some advantages and disadvantages of FIR filters are sum-
marized as follows:
1. Because there is no feedback of past outputs as defined in (3.1.16), the FIR filters
are always stable. That is, a bounded input results in a bounded output. This
inherent stability is also manifested in the absence of poles in the transfer function
as defined in (4.3.8), except possibly at the origin.
2. The filter has finite memory because it `forgets' all inputs before the (L ÷1)th
previous one.
3. The design of linear phase filters can be guaranteed. In applications such as audio
signal processing and data transmission, linear phase filters are preferred since they
avoid phase distortion.
4. The finite-precision errors (discussed in Chapter 3) are less severe in FIR filters than
in IIR filters.
5. FIR filters can be easily implemented on most DSP processors such as the
TMS320C55x introduced in Chapter 2.
6. A relatively higher order FIR filter is required to obtain the same characteristics as
compared with an IIR filter. Thus more computations are required, and/or longer
time delay may be involved in the case of FIR filters.
5.1.2 Filter Types
An ideal frequency-selective filter passes certain frequency components without any
change and completely stops the other frequencies. The range of frequencies that are
INTRODUCTION TO DIGITAL FILTERS 183
passed without attenuation is the passband of the filter, and the range of frequencies
that is attenuated is the stopband. Thus the magnitude response of an ideal filter is given
by [H(!)[ = 1 in the passband and [H(!)[ = 0 in the stopband. Note that the frequency
response H(!) of a digital filter is a periodic function of !, and the magnitude response
[H(!)[ of a digital filter with real coefficients is an even function of !. Therefore the
digital filter specifications are given only for the range 0 _ ! _ p.
The magnitude response of an ideal lowpass filter is illustrated in Figure 5.1(a). The
regions 0 _ ! _ !
c
and ! > !
c
are referred to as the passband and stopband, respec-
tively. The frequency that separates the passband and stopband is called the cut-off
frequency !
c
. An ideal lowpass filter has magnitude response [H(!)[ = 1 in the fre-
quency range 0 _ ! _ !
c
and has [H(!)[ = 0 for ! > !
c
. Thus a lowpass filter passes all
low-frequency components below the cut-off frequency and attenuates all high-fre-
quency components above !
c
. Lowpass filters are generally used when the signal
components of interest are in the range of DC to the cut-off frequency, but other higher
frequency components (or noise) are present.
The magnitude response of an ideal highpass filter is illustrated in Figure 5.1(b). The
regions ! _ !
c
and 0 _ ! < !
c
are referred to as the passband and stopband, respec-
tively. A highpass filter passes all high-frequency components above the cut-off fre-
quency !
c
and attenuates all low-frequency components below !
c
. As discussed in
Chapter 1, highpass filters can be used to eliminate DC offset, 60 Hz hum, and other
low frequency noises.
The magnitude response of an ideal bandpass filter is illustrated in Figure 5.1(c). The
regions ! < !
a
and ! > !
b
are referred to as the stopband. The frequencies !
a
and !
b
are called the lower cut-off frequency and the upper cut-off frequency, respectively. The
H(w) H(w)
H(w) H(w)
w
1
0 w
c
p
0 w
a
w
b
p 0 w
a
w
b
p
0 w
c
p
w
1
(a) (b)
w w
1 1
(c) (d)
Figure 5.1 Magnitude response of ideal filters: (a) lowpass, (b) highpass, (c) bandpass, and
(d) bandstop
184 DESIGN AND IMPLEMENTATION OF FIR FILTERS
region !
a
_ ! _ !
b
is called the passband. A bandpass filter passes all frequency
components between the two cut-off frequencies !
a
and !
b
, and attenuates all fre-
quency components below the frequency !
a
and above the frequency !
b
. If the passband
is narrow, it is more common to specify the center frequency and the bandwidth of the
passband. A narrow bandpass filter may be called a resonator (or peaking) filter.
The magnitude response of an ideal bandstop (or band-reject) filter is illustrated
in Figure 5.1(d). The regions ! _ !
a
and ! _ !
b
are referred to as the passband.
The region !
a
< ! < !
b
is called the stopband. A bandstop filter attenuates all fre-
quency components between the two cutoff frequencies !
a
and !
b
, and passes all
frequency components below the frequency !
a
and above the frequency !
b
. A narrow
bandstop filter designed to attenuate a single frequency component is called a notch
filter. For example, a common source of noise is a power line generating a 60 Hz
sinusoidal signal. This noise can be removed by passing the corrupted signal through
a notch filter with notch frequency at 60 Hz. The design of simple notch filters was
introduced in Section 4.5.1.
In addition to these frequency-selective filters, an allpass filter provides frequency
response [H(!)[ = 1 for all frequency !, thus passing all frequencies with uniform gain.
These filters do not remove frequency components, but alter the phase response. The
principal use of allpass filters is to correct the phase distortion introduced by the
physical system and/or other filters. For example, it is used as a delay equalizer. In
this application, it is designed such that when cascaded with another digital system, the
overall system has a constant group delay in the frequency range of interest. A very
special case of the allpass filter is the ideal Hilbert transformer, which produces a 908
phase shift of input signals.
5.1.3 Filter Specifications
In practice, we cannot achieve the infinitely sharp cutoff implied by the ideal filters
shown in Figure 5.1. This will be shown later by considering the impulse response of the
ideal lowpass filter that is non-causal and hence not physically realizable. Instead we
must compromise and accept a more gradual cutoff between passband and stopband, as
well as specify a transition band between the passband and stopband. The design is
based on magnitude response specifications only, so the phase response of the filter is
not controlled. Whether this is important depends on the application. Realizable filters
do not exhibit the flat passband or the perfect linear phase characteristic. The deviation
of [H(!)[ from unity (0 dB) in the passband is called magnitude distortion, and the
deviation from the linear phase of the phase response H(!) is called phase distortion.
The characteristics of digital filters are often specified in the frequency domain. For
frequency-selective filters, the magnitude response specifications of a digital filter are
often given in the form of tolerance (or ripple) schemes. In addition, a transition band is
specified between the passband and the stopband to permit the magnitude drop off
smoothly. A typical magnitude response of lowpass filter is shown in Figure 5.2. The
dotted horizontal lines in the figure indicate the tolerance limits. In the passband, the
magnitude response has a peak deviation d
p
and in the stopband, it has a maximum
deviation d
s
. The frequencies !
p
and !
s
are the passband edge (cut-off) frequency and
the stopband edge frequency, respectively.
INTRODUCTION TO DIGITAL FILTERS 185
Passband
A
p
A
s
Stopband
1+d
p
1−d
p
0 w
p
w
c
w
s
d
s
H(w)
w
p
Ideal filter
Actual filter
Transition
band
1
Figure 5.2 Magnitude response and performance measurement of a lowpass filter
As shown in Figure 5.2, the magnitude of passband defined by 0 _ ! _ !
p
approxi-
mates unity with an error of ±d
p
. That is,
1 ÷d
p
_ [H(!)[ _ 1 ÷d
p
, 0 _ ! _ !
p
: (5:1:4)
The passband ripple, d
p
, is a measure of the allowed variation in magnitude response in
the passband of the filter. Note that the gain of the magnitude response is normalized to
1 (0 dB). In practical applications, it is easy to scale the filter output by multiplying the
output by a constant, which is equivalent to multiplying the whole magnitude response
by the same constant gain.
In the stopband, the magnitude approximates 0 with an error d
s
. That is,
[H(!)[ _ d
s
, !
s
_ ! _ p: (5:1:5)
The stopband ripple (or attenuation) describes the maximum gain (or minimum
attenuation) for signal components above the !
s
.
Passband and stopband deviations may be expressed in decibels. The peak passband
ripple, d
p
, and the minimum stopband attenuation, d
s
, in decibels are given as
A
p
= 20 log
10
1 ÷d
p
1 ÷d
p

dB (5:1:6)
and
A
s
= ÷20 log
10
d
s
dB: (5:1:7)
Thus we have
186 DESIGN AND IMPLEMENTATION OF FIR FILTERS
d
p
=
10
A
p
=20
÷1
10
A
p
=20
÷1
(5:1:8)
and
d
s
= 10
÷A
s
=20
: (5:1:9)
Example 5.1: Consider a filter specified as having a magnitude response in the
passband within ±0:01. That is, d
p
= 0:01. From (5.1.6), we have
A
p
= 20 log
10
1:01
0:99

= 0:1737 dB:
When the minimum stopband attenuation is given as d
s
= 0:01, we have
A
s
= ÷20 log
10
(0:01) = 40 dB:
The transition band is the area between the passband edge frequency !
p
and the
stopband edge frequency !
s
. The magnitude response decreases monotonically from the
passband to the stopband in this region. Generally, the magnitude in the transition band
is left unspecified. The width of the transition band determines how sharp the filter is. It
is possible to design filters that have minimum ripple over the passband, but a certain
level of ripple in this region is commonly accepted in exchange for a faster roll-off of
gain in the transition band. The stopband is chosen by the design specifications.
Generally, the smaller d
p
and d
s
are, and the narrower the transition band, the more
complicated (higher order) the designed filter becomes.
An example of a narrow bandpass filter is illustrated in Figure 5.3. The center
frequency !
m
is the point of maximum gain (or maximum attenuation for a notch
filter). If a logarithm scale is used for frequency such as in many audio applications, the
center frequency at the geometric mean is expressed as
!
m
=

!
a
!
b

, (5:1:10a)
H(w)
w
1
2
1
w
a
w
m
w
b
Figure 5.3 Magnitude response of bandpass filter with narrow bandwidth
INTRODUCTION TO DIGITAL FILTERS 187
where !
a
and !
b
are the lower and upper cut-off frequencies, respectively. The
bandwidth is the difference between the two cut-off frequencies for a bandpass filter.
That is,
BW= !
b
÷!
a
: (5:1:10b)
The 3-dB bandwidth commonly used in practice is defined as
[H(!
a
)[ = [H(!
b
)[ =
1

2
÷ 0:707: (5:1:11)
Another way of describing a resonator (or notch) filter is the quality factor defined as
Q =
!
m
2pBW
: (5:1:12)
There are many applications that require high Q filters.
When a signal passes through a filter, it is modified both in amplitude and phase. The
phase response is an important filter characteristic because it affects time delay of the
different frequency components passing through the filter. If we consider a signal that
consists of several frequency components, the phase delay of the filter is the average
time delay the composite signal suffers at each frequency. The group delay function is
defined as
T
d
(!) =
÷df(!)
d!
, (5:1:13)
where f(!) is the phase response of the filter.
A filter is said to have a linear phase if its phase response satisfies
f(!) = ÷a!, ÷p _ ! _ p (5:1:14)
or
f(!) = b ÷a!, ÷p _ ! _ p (5:1:15)
These equations show that for a filter with a linear phase, the group delay T
d
(!) given in
(5.1.13) is a constant a for all frequencies. This filter avoids phase distortion because all
sinusoidal components in the input are delayed by the same amount. A filter with a
nonlinear phase will cause a phase distortion in the signal that passes through it. This is
because the frequency components in the signal will each be delayed by a different
amount, thereby altering their harmonic relationships. Linear phase is important in data
communications, audio, and other applications where the temporal relationships
between different frequency components are critical.
The specifications on the magnitude and phase (or group delay) of H(!) are based on
the steady-state response of the filter. Therefore they are called the steady-state speci-
fications. The speed of the response concerns the rate at which the filter reaches the
steady-state response. The transient performance is defined for the response right after
188 DESIGN AND IMPLEMENTATION OF FIR FILTERS
the application of an input signal. A well-designed filter should have a fast response, a
small rise time, a small settling time, and a small overshoot.
In theory, both the steady-state and transient performance should be considered in
the design of a digital filter. However, it is difficult to consider these two specifications
simultaneously. In practice, we first design a filter to meet the magnitude specifications.
Once this filter is obtained, we check its phase response and transient performance. If
they are satisfactory, the design is completed. Otherwise, we must repeat the design
process. Once the transfer function has been determined, we can obtain a realization of
the filter. This will be discussed later.
5.2 FIR Filtering
The signal-flow diagram of the FIR filter is shown in Figure 3.6. As discussed in
Chapter 3, the general I/O difference equation of FIR filter is expressed as
y(n) = b
0
x(n) ÷b
1
x(n ÷1) ÷ ÷b
L÷1
x(n ÷L ÷1) =
¸
L÷1
l=0
b
l
x(n ÷l), (5:2:1)
where b
l
are the impulse response coefficients of the FIR filter. This equation describes
the output of the FIR filter as a convolution sum of the input with the impulse response
of the system. The transfer function of the FIR filter defined in (5.2.1) is given by
H(z) = b
0
÷b
1
z
÷1
÷ ÷b
L÷1
z
÷(L÷1)
=
¸
L÷1
l=0
b
l
z
÷l
: (5:2:2)
5.2.1 Linear Convolution
As discussed in Section 3.2.2, the output of the linear system defined by the impulse
response h(n) for an input signal x(n) can be expressed as
y(n) = x(n) + h(n) =
¸
·
l=÷·
h(l)x(n ÷l): (5:2:3)
Thus the output of the LTI system at any given time is the sum of the input samples
convoluted by the impulse response coefficients of the system. The output at time n
0
is
given as
y(n
0
) =
¸
·
l=÷·
h(l)x(n
0
÷l): (5:2:4)
Assuming that n
0
is positive, the process of computing the linear convolution involves
the following four steps:
FIR FILTERING 189
1. Folding. Fold x(l) about l = 0 to obtain x(÷l).
2. Shifting. Shift x(÷l) by n
0
samples to the right to obtain x(n
0
÷l).
3. Multiplication. Multiply h(l) by x(n
0
÷l) to obtain the products h(l) x(n
0
÷l) for
all l.
4. Summation. Sum all the products to obtain the output y(n
0
) at time n
0
.
Repeat steps 2±4 in computing the output of the system at other time instants n
0
.
This general procedure of computing convolution sums can be applied to (5.2.1) for
calculating the FIR filter output y(n). As defined in (3.2.15), the impulse response of the
FIR filter is
h(l) =
0, l < 0
b
l
, 0 _ l < L
0, l _ L.

(5:2:5)
If the input signal is causal, the general linear convolution equation defined in (5.2.3)
can be simplified to (5.2.1). Note that the convolution of the length M input with the
length L impulse response results in length L ÷M÷1 output.
Example 5.2: Consider an FIR filter that consists of four coefficients b
0
, b
1
, b
2
,
and b
3
. From (5.2.1), we have
y(n) =
¸
3
l=0
b
l
x(n ÷l), n _ 0:
This yields
n = 0, y(0) = b
0
x(0),
n = 1, y(1) = b
0
x(1) ÷b
1
x(0),
n = 2, y(2) = b
0
x(2) ÷b
1
x(1) ÷b
2
x(0),
n = 3, y(3) = b
0
x(3) ÷b
1
x(2) ÷b
2
x(1) ÷b
3
x(0):
In general, we have
y(n) = b
0
x(n) ÷b
1
x(n ÷1) ÷b
2
x(n ÷2) ÷b
3
x(n ÷3), n _ 3:
The graphical interpretation is illustrated in Figure 5.4.
As shown in Figure 5.4, the input sequence is flipped around (folding) and then
shifted to the right over the filter coefficients. At each time instant, the output value is
the sum of products of overlapped coefficients with the corresponding input data
190 DESIGN AND IMPLEMENTATION OF FIR FILTERS
aligned below it. This flip-and-slide form of linear convolution can be illustrated in
Figure 5.5. Note that shifting x(÷l) to the right is equivalent to shift b
l
to the left one
unit at each sampling period.
As shown in Figure 5.5, the input sequence is extended by padding L ÷1 zeros to its
right. At time n = 0, the only non-zero product comes from b
0
and x(0) which are time
aligned. It takes the filter L ÷1 iterations before it is completely overlapped with the
input sequence. The first L ÷1 outputs correspond to the transient behavior of the FIR
filter. For n _ L ÷1, the filter aligns over the non-zero portion of the input sequence.
That is, the signal buffer of FIR filter is full and the filter is in the steady state. If the
input is a finite-length sequence of M samples, there are L ÷M÷1 output samples and
the last L ÷1 samples also correspond to transients.
x(n−1)
x(n−2)
x(n−3)
b
0
x(0)
b
0
x(1)
b
1
x(1)
b
1
x(n−1)
b
0
x(2)
b
0
x(n)
b
1
x(0)
b
2
x(0)
b
2
x(n−2)
b
3
x(n−3)
b
0
b
1
b
2
b
3
n = 0:
n = 1:
n = 2:
n ≥ 3:
x(0)
x(0)
x(0)
x(1)
x(1)
x(2)
x(n)
Figure 5.4 Graphical interpretation of linear convolution, L = 4
b
0
b
0
b
1
b
1
b
2
b
2
b
3
b
3
x(n) x(1)
y(n) y(0)
x(n−1) x(n−2) x(n−3) 0 0 0
Figure 5.5 Flip-and-slide process of linear convolution
FIR FILTERING 191
5.2.2 Some Simple FIR Filters
A multiband filter has more than one passband and stopband. A special case of the
multiband filter is the comb filter. A comb filter has evenly spaced zeros, with the shape
of the magnitude response resembling a comb in order to block frequencies that are
integral multiples of a fundamental frequency. A difference equation of a comb filter is
given as
y(n) = x(n) ÷x(n ÷L), (5:2:6)
where the number of delay L is an integer. The transfer function of this multiplier-free
FIR filter is
H(z) = 1 ÷z
÷L
=
z
L
÷1
z
L
: (5:2:7)
Thus the comb filter has L poles at the origin (trivial poles) and L zeros equally spaced
on the unit circle at
z
l
= e
j(2p=L)l
, l = 0, 1, . . . , L ÷1: (5:2:8)
Example 5.3: The zeros and the frequency response of a comb filter can be
computed and plotted using the following MATLAB script for L = 8:
b = [1 0 0 0 0 0 0 0 ÷1];
zplane(b, 1)
freqz(b, 1, 128);
The zeros on the z-plane are shown in Figure 5.6(a) and the characteristic of comb
shape is shown in Figure 5.6(b). The center of the passband lies halfway between
the zeros of the response, that is at frequencies
(2l ÷1)p
L
, l = 0, 1, . . . , L ÷1.
Because there is not a large attenuation in the stopband, the comb filter can
only be used as a crude bandstop filter to remove harmonics at frequencies
!
l
= 2pl=L, l = 0, 1, . . . , L ÷1: (5:2:9)
Comb filters are useful for passing or eliminating specific frequencies and their
harmonics. Periodic signals have harmonics and using comb filters are more efficient
than having individual filters for each harmonic. For example, the constant humming
sound produced by large transformers located in electric utility substations are com-
posed of even-numbered harmonics (120 Hz, 240 Hz, 360 Hz, etc.) of the 60 Hz power
frequency. When a desired signal is corrupted by the transformer noise, the comb filter
with notches at the multiples of 120 Hz can be used to eliminate undesired harmonics.
We can selectively cancel one or more zeros in a comb filter with corresponding poles.
Canceling the zero provides a passband, while the remaining zeros provide attenuation
for a stopband. For example, we can add a pole at z = 1. Thus the transfer function
given in (5.2.7) is changed to
192 DESIGN AND IMPLEMENTATION OF FIR FILTERS
1
0.8
0.6
0.4
0.2
0 (a)
(b)
−0.2
−0.4
−0.6
−0.8
−1
−1 −0.5 0
Real Part
I
m
a
g
i
n
a
r
y

P
a
r
t
0.5 1
0
10
5
0
−5
−10
−15
0.1 0.2 0.3 0.4 0.5
M
a
g
n
i
t
u
d
e

(
d
B
)
Normalized Frequency (ϫp rad/sample)
0.6 0.7 0.8 0.9 1
0
100
50
0
−50
−100
0.1 0.2 0.3 0.4 0.5
P
h
a
s
e

(
d
e
g
r
e
e
s
)
Normalized Frequency (ϫp rad/sample)
0.6 0.7 0.8 0.9 1
Figure 5.6 Zeros of a simple comb filter (L = 8) and its frequency response: (a) zeros, and
(b) magnitude (top) and phase (bottom) responses
H(z) =
1 ÷z
÷L
1 ÷z
÷1
: (5:2:10)
This is a lowpass filter with passband centered at z = 1, where the pole±zero cancella-
tion occurs. Since the pole at z = 1 is canceled by the zero at z = 1, the system defined
by (5.2.10) is still the FIR filter. Note that canceling the zero at z = 1 produces a
lowpass filter, canceling the zeros at z = ±j produces a bandpass filter, and canceling
the zero at z = ÷1 produces a highpass filter.
Applying the scaling factor 1/L to (5.2.10), the transfer function becomes
H(z) =
1
L
1 ÷z
÷L
1 ÷z
÷1

: (5:2:11)
This is the moving-average filter introduced in Chapter 3 with the I/O difference
equation expressed as
FIR FILTERING 193
y(n) =
1
L
[x(n) ÷x(n ÷L) ÷y(n ÷1)[
=
1
L
¸
L÷1
l=0
x(n ÷l): (5:2:12)
The moving-average filter is a very simple lowpass filtering operation that passes the
zero-frequency (or the mean) component. However, there are disadvantages of this
type of filter such as the passband cut-off frequency is a function of L and the sampling
rate f
s
, and the stopband attenuation is fixed by L.
Example 5.4: Consider a simple moving-average filter
y(n) =
1
2
[x(n) ÷x(n ÷1)[, n _ 0:
The transfer function of the filter can be expressed as
H(z) =
1
2
1 ÷z
÷1

,
which has a single zero at z = ÷1 and the frequency response is given by
H(!) =
1
2
1 ÷e
÷j!

=
1
2
e
÷j!=2
e
j!=2
÷e
÷j!=2

= e
÷j!=2
cos(!=2):
Therefore, we have
[H(!)[
2
= cos
!
2

2
=
1
2
[1 ÷cos(!)[:
This is lowpass-filter response, which falls off monotonically to 0 at ! = p. We can
show that
f(!) =
÷!
2
,
thus the filter has linear phase.
5.2.3 Linear Phase FIR Filters
In many practical applications, it is required that a digital filter has a linear phase. In
particular, it is important for phase-sensitive signals such as speech, music, images, and
data transmission where nonlinear phase would give unacceptable frequency distortion.
FIR filters can be designed to obtain exact linear phase.
If L is an odd number, we define M = (L ÷1)=2. If we define h
l
= b
l÷M
, then (5.2.1)
can be written as
194 DESIGN AND IMPLEMENTATION OF FIR FILTERS
B(z) =
¸
2M
l=0
b
l
z
÷l
=
¸
M
l=÷M
b
l÷M
z
÷(l÷M)
= z
÷M
¸
M
l=÷M
h
l
z
÷l
¸ ¸
= z
÷M
H(z), (5:2:13)
where
H(z) =
¸
M
l=÷M
h
l
z
÷l
: (5:2:14)
Let h
l
have the symmetry property expressed as
h
l
= h
÷l
, l = 0, 1, . . . , M: (5:2:15)
From (5.2.13), the frequency response B(!) can be written as
B(!) = B(z)[
z = e
j!
= e
÷j!M
H(!)
= e
÷j!M
¸
M
l=÷M
h
l
e
÷j!l
¸ ¸
= e
÷j!M
h
0
÷
¸
M
l=1
h
l
e
j!l
÷e
÷j!l

¸ ¸
= e
÷j!M
h
0
÷2
¸
M
l=1
h
l
cos(!l)
¸ ¸
:
(5:2:16)
If h
l
is real, then H(!) is a real function of !. If H(!) _ 0, then the phase of B(!) is
equal to
f(!) = ÷!M, (5:2:17)
which is a linear function of !. However, if H(!) < 0, then the phase of B(!) is equal to
p ÷!M. Thus, if there are sign changes in H(!), there are corresponding 1808 phase
shifts in B(!), and B(!) is only piecewise linear. However, it is still simple to refer to the
filter as having linear phase.
If h
l
has the anti-symmetry property expressed as
h
l
= ÷h
÷l
, l = 0, 1, . . . , M, (5:2:18)
this implies h(0) = 0. Following the derivation of (5.2.16), we obtain
B(!) = e
÷j!M
H(!) = e
÷j!M
¸
M
l=1
h
l
e
j!l
÷e
÷j!l

¸ ¸
= e
÷j!M
÷2j
¸
M
l=1
h
l
sin(!l)
¸ ¸
:
(5:2:19)
If h
l
is real, then H(!) is pure imaginary and the phase of B(z) is a linear function of !.
The filter order L is assumed to be an odd integer in the above derivations. If L is an
even integer and M = L=2, then the derivations of (5.2.16) and (5.2.19) have to be
FIR FILTERING 195
modified slightly. In conclusion, an FIR filter has linear phase if its coefficients satisfy
the following (positive) symmetric condition:
b
l
= b
L÷1÷l
, l = 0, 1, . . . , L ÷1, (5:2:20)
or, the anti-symmetric (negative symmetry) condition
b
l
= ÷b
L÷1÷l
, l = 0, 1, . . . , L ÷1: (5:2:21)
There are four types of linear phase FIR filters, depending on whether L is even or
odd and whether b
l
has positive or negative symmetry as illustrated in Figure 5.7. The
group delay of a symmetric (or anti-symmetric) FIR filter is T
d
(!) = L=2, which
corresponds to the midpoint of the FIR filter. The frequency response of the type I
l
Center of
symmetry
(a)
l
(b)
l
(c)
l
(d)
Center of
symmetry
Center of
symmetry
Center of
symmetry
Figure 5.7 Coefficients of the four types of linear phase FIR filters: (a) type I: L even (L = 8),
positive symmetry, (b) type II: L odd (L = 7), positive symmetry, (c) type III: L even (L = 8),
negative symmetry, and (d) type IV: L odd (L = 7), negative symmetry
196 DESIGN AND IMPLEMENTATION OF FIR FILTERS
(L even, positive symmetry) filter is always 0 at the Nyquist frequency. This type of filter
is unsuitable for a highpass filter. Type III (L even, negative symmetry) and IV (L odd,
negative symmetry) filters introduce a 908 phase shift, thus they are often used to design
Hilbert transformers. The frequency response is always 0 at DC frequency, making
them unsuitable for lowpass filters. In addition, type III response is always 0 at the
Nyquist frequency, also making it unsuitable for a highpass filter.
The symmetry (or anti-symmetry) property of a linear-phase FIR filter can be
exploited to reduce the total number of multiplications into almost half. Consider the
realization of FIR filter with an even length L and positive symmetric impulse response
as given in (5.2.20), Equation (5.2.2) can be combined as
H(z) = b
0
1 ÷z
÷L÷1

÷b
1
z
÷1
÷z
÷L÷2

÷ ÷b
L=2÷1
z
÷L=2÷1
÷z
÷L=2

: (5:2:22)
The I/O difference equation is given as
y(n) = b
0
[x(n) ÷x(n ÷L ÷1)[ ÷b
1
[x(n ÷1) ÷x(n ÷L ÷2)[
÷ ÷b
L=2÷1
[x(n ÷L=2 ÷1) ÷x(n ÷L=2)[
=
¸
L=2÷1
l=0
b
l
[x(n ÷l) ÷x(n ÷L ÷1 ÷l)[: (5:2:23)
A realization of H(z) defined in (5.2.22) is illustrated in Figure 5.8. For an anti-
symmetric FIR filter, the addition of two signals is replaced by subtraction. That is,
y(n) =
¸
L=2÷1
l=0
b
l
[x(n ÷l) ÷x(n ÷L ÷1 ÷l)[: (5:2:24)
As shown in (5.2.23) and Figure 5.8, the number of multiplications is cut in half by
adding the pair of samples, then multiplying the sum by the corresponding coefficient.
x(n)
y(n)
x(n−1) x(n−L/2+1)
x(n−L/2)
x(n−L+1)
x(n−L+2)
z
−1
z
−1
z
−1
z
−1
z
−1
z
−1
z
−1
b
0
b
1
b
L/2−1
Figure 5.8 Signal flow diagram of symmetric FIR filter, L is even
FIR FILTERING 197
The trade-off is that instead of accessing data linearly through the same buffer with a
single pointer, we need two address pointers that point at both ends for x(n ÷l) and
x(n ÷L ÷1 ÷l). The TMS320C55x provides two special instructions for implementing
the symmetric and anti-symmetric FIR filters efficiently. In Section 5.6, we will demon-
strate how to use the symmetric FIR instructions for experiments.
There are applications where data is already collected and stored for later processing,
i.e., the processing is not done in real time. In these cases, the `current' time n can be
located arbitrarily as the data is processed, so that the current output of the filter may
depend on past, current, and future input values. Such a filter is `non-realizable' in real
time, but is easy to implement for the stored data. The non-causal filter has the I/O
equation
y(n) =
¸
L
2
l=÷L
1
b
l
x(n ÷l) (5:2:25)
and the transfer function
H(z) =
¸
L
2
l=÷L
1
b
l
z
÷l
: (5:2:26)
Some typical applications of non-causal filters are the smoothing filters, the interpola-
tion filters, and the inverse filters. A simple example of a non-causal filter is a Hanning
filter with coefficients ¦0:25, 0:5, 0:25¦ for smoothing estimated pitch in speech pro-
cessing.
5.2.4 Realization of FIR Filters
An FIR filter can be realized to operate either on a block basis or a sample-by-sample
basis. In the block processing case, the input is segmented into multiple blocks. Filtering
is performed on one block at a time, and the resulting output blocks are recombined to
form the overall output. The filtering of each block can be implemented using the linear
convolution technique discussed in Section 5.2.1, or fast convolution using FFT, which
will be introduced in Chapter 7. The implementation of block-FIR filter with the
TMS320C55x will be introduced in Section 5.6. In the sample-by-sample case, the
input samples are processed at every sampling period after the current input x(n) is
available. This approach is useful in real-time applications and will be discussed in this
section.
Once the coefficients, b
l
, l = 0, 1, . . . , L ÷1, have been determined, the next step is
to decide on the structure (form) of the filter. The direct form implementation of
digital FIR (transversal) filter is shown in Figure 3.6 and is described by the difference
equation (5.2.1). The transfer function of the FIR filter given in (5.2.2) can be factored
as
H(z) = b
0
(1 ÷z
1
z
÷1
)(1 ÷z
2
z
÷1
) (1 ÷z
L÷1
z
÷1
), (5:2:27)
198 DESIGN AND IMPLEMENTATION OF FIR FILTERS
where the zeros z
l
, l = 1, 2, , L ÷1 must occur in complex-conjugate pairs for a real-
valued filter. The factorization of H(z) can be carried out in MATLAB using the
function roots. If we pair two zeros and multiply out the expressions, we obtain a
cascade of second-order sections as
H(z) = b
0
(1 ÷b
11
z
÷1
÷b
12
z
÷2
)(1 ÷b
21
z
÷1
÷b
22
z
÷2
) (1 ÷b
M1
z
÷1
÷b
M2
z
÷2
)
= b
0
¸
M
m=1
(1 ÷b
m1
z
÷1
÷b
m2
z
÷2
)
= b
0
H
1
(z)H
2
(z) H
M
(z),
(5:2:28)
where M = (L ÷1)=2 if L is odd and M = L=2 if L is even. Thus the higher order H(z)
given in (5.2.2) is broken up and can be implemented in cascade form as illustrated in
Figure 5.9. Splitting the filter in this manner reduces roundoff errors, which may be
critical for some applications. However, the direct form is more efficient for implemen-
tation on most commercially available DSP processors such as the TMS320C55x.
The output y(n) is a linear combination of a finite number of inputs ¦x(n), x(n ÷1),
. . . , x(n ÷L ÷1)¦ and L coefficients ¦b
l
, l = 0, 1, . . . , L ÷1¦, which can be represented
as tables illustrated in Figure 5.10. In order to compute the output at any time, we
simply have to multiply the corresponding values in each table and sum the results.
That is,
y(n) = b
0
x(n) ÷b
1
x(n ÷1) ÷ ÷b
L÷1
x(n ÷L ÷1): (5:2:29)
In FIR filtering, the coefficient values are constant, but the data in the signal buffer
changes every sampling period, T. That is, the x(n) value at time n becomes x(n ÷1) in
the next sampling period, then x(n ÷2), etc., until it simply drops off the end of the
delay chain.
x(n) y(n)
b
0
H
1
(z) H
2
(z) H
M
(z)
Input
Output
b
m2
b
m1
z
−1
z
−1
(a)
(b)
Figure 5.9 A cascade structure of FIR filter: (a) overall structure, and (b) flow diagram of
second-order FIR section
FIR FILTERING 199
x (n − L + 1)
Coefficients buffer
b
0
b
1
b
2
x(n)
x(n − 1)
x(n − 2)
b
L−1
Signal buffer
Figure 5.10 Tables of coefficient vector and signal buffer
x(n) x(n − 1)
x(n) x(n − 1)
x(n − 2)
x(n − 2)
. . .
x(n − L + 2)x(n − L + 1)
. . .
x(n − L + 1)
Current
Time, n
Next
time, n+1
Discarded
Next data
Figure 5.11 Refreshing the signal buffer for FIR filtering
The signal buffer is refreshed in every sampling period in the fashion illustrated in
Figure 5.11, where the oldest sample x(n ÷L ÷1) is discarded and other signals are
shifted one location to the right in the buffer. A new sample (from ADC in real-time
application) is inserted to the memory location labeled as x(n). The FIR filtering
operation that computes y(n) using (5.2.29) is then performed. The process of refreshing
the signal buffer shown in Figure 5.11 requires intensive processing time if the operation
is not implemented by the DSP hardware.
The most efficient method for handling a signal buffer is to load the signal samples
into a circular buffer, as illustrated in Figure 5.12(a). Instead of shifting the data
forward while holding the buffer addresses fixed, the data is kept fixed and the addresses
are shifted backwards (counterclockwise) in the circular buffer. The beginning of the
signal sample, x(n), is pointed at with a pointer and the previous samples are loaded
sequentially from that point in a clockwise direction. As we receive a new sample, it is
placed at the position x(n) and our filtering operation defined in (5.2.29) is performed.
After calculating the output y(n), the pointer is moved counterclockwise one position to
the point at x(n ÷L ÷1) and we wait for the next input signal. The next input at time
n ÷1 is written to the x(n ÷L ÷1) position, and is referred to as x(n) for the next
iteration. This is permitted because the old x(n ÷L ÷1) signal dropped off the end of
our delay chain after the previous calculation as shown in Figure 5.11. The circular
buffer implementation of a signal buffer, or a tapped-delay-line is very efficient. The
update is carried out by adjusting the address pointer without physically shifting any
data in memory. It is especially useful in implementing a comb filter when L is large,
since we only need to access two adjacent samples x(n) and x(n ÷L) in the circular
200 DESIGN AND IMPLEMENTATION OF FIR FILTERS
x(n − L + 2)
x(n − L + 1)
x(n − 1)
x(n)
x(n − 2)
x(n − 3)
Signal buffer pointer
at time n
Signal buffer pointer
for next x(n)
(a)
b
0
b
1
b
2
b
3
b
L−1
b
L−2
Coefficient
buffer pointer
(b)
Figure 5.12 Circular buffers for FIR filter: (a) circular buffer for holding the signals for FIR
filtering. The pointer to x(n) is updated in the counterclockwise direction, and (b) circular buffer for
FIR filter coefficients, the pointer always pointing to b
0
at the beginning of filtering
buffer. It is also used in sinewave generators and wavetable sound synthesis, where a
stored waveform can be generated periodically by cycling over the circular buffer.
Figure 5.12(b) shows a circular buffer for FIR filter coefficients. Circular buffer
allows the coefficient pointer to wrap around when it reaches to the end of the
coefficient buffer. That is, the pointer moves from b
L÷1
to b
0
such that the filtering
will always start with the first coefficient.
5.3 Design of FIR Filters
The objective of FIR filter design is to determine a set of filter coefficients ¦b
l
,
l = 0, 1, . . . , L ÷1¦ such that the filter performance is close to the given specifications.
A variety of techniques have been proposed for the design of FIR filters. In this section,
we discuss two direct design methods. The first method is based on truncating the
Fourier series representation of the desired frequency response. The second method is
based on specifying equally spaced frequency samples of the frequency response of the
desired filter.
5.3.1 Filter Design Procedure
The design of digital FIR filters involves five steps:
1. Specification of filter requirements.
2. Calculation and optimization of filter coefficients.
3. Realization of the filter by a suitable structure.
DESIGN OF FIR FILTERS 201
4. Analysis of finite wordlength effects on filter performance.
5. Implementation of filter in software and/or hardware.
These five steps are not necessarily independent, and they may be conducted in a
different order. Specification of filter characteristics and realization of desired filters
were discussed in Section 5.2. In this section, we focus on designing FIR filters for given
specifications.
There are several methods for designing FIR filters. The methods discussed in this
section are the Fourier series (window) method and the frequency-sampling method.
The Fourier series method offers a very simple and flexible way of computing FIR filter
coefficients, but it does not allow the designer adequate control over the filter para-
meters. The main attraction of the frequency-sampling method is that it allows recursive
realization of FIR filters, which can be computationally efficient. However, it lacks
flexibility in specifying or controlling filter parameters.
With the availability of an efficient and easy-to-use filter design program such as
MATLAB, the Park±McClellan algorithm is now widely used in industry for FIR filter
design. The Park±McClellan algorithm should be the method of first choice for most
practical applications.
5.3.2 Fourier Series Method
The basic idea of Fourier series method is to design an FIR filter that approximates the
desired frequency response of filter by calculating its impulse response. This method
utilizes the fact that the frequency response H(!) of a digital filter is a periodic function
with period 2p. Thus it can be expanded in a Fourier series as
H(!) =
¸
·
n=÷·
h(n)e
÷j!n
, (5:3:1)
where
h(n) =
1
2p

p
÷p
H(!)e
j!n
d!, ÷·_ n _ ·: (5:3:2)
This equation shows that the impulse response h(n) is double-sided and has infinite
length. If H(!) is an even function in the interval [![ _ p, we can show that (see exercise
problem)
h(n) =
1
p

p
0
H(!) cos(!n)d!, n _ 0 (5:3:3)
and the impulse response is symmetric about n = 0. That is,
h(÷n) = h(n), n _ 0: (5:3:4)
202 DESIGN AND IMPLEMENTATION OF FIR FILTERS
For a given desired frequency response H(!), the corresponding impulse response
(filter coefficients) h(n) can be calculated for a non-recursive filter if the integral (5.3.2)
or (5.3.3) can be evaluated. However, in practice there are two problems with this simple
design technique. First, the impulse response for a filter with any sharpness to its
frequency response is infinitely long. Working with an infinite number of coefficients
is not practical. Second, with negative values of n, the resulting filter is non-causal, thus
is non-realizable for real-time applications.
A finite-duration impulse response ¦h
/
(n)¦ of length L = 2M÷1 that is the best
approximation (minimum mean-square error) to the ideal infinite-length impulse
response can be simply obtained by truncation. That is,
h
/
(n) =
h(n), ÷M _ n _ M
0, otherwise.

(5:3:5)
Note that in this definition, we assume L to be an odd number otherwise M will not be
an integer. On the unit circle, we have z = e
j!
and the system transfer function is
expressed as
H
/
(z) =
¸
M
n=÷M
h
/
(n)z
÷n
: (5:3:6)
It is clear that this filter is not physically realizable in real time since the filter must
produce an output that is advanced in time with respect to the input.
A causal FIR filter can be derived by delaying the h
/
(n) sequence by M samples.
That is, by shifting the time origin to the left of the vector and re-indexing the
coefficients as
b
/
l
= h
/
(l ÷M), l = 0, 1, . . . , 2M: (5:3:7)
The transfer function of this causal FIR filter is
B
/
(z) =
¸
L÷1
l=0
b
/
l
z
÷l
: (5:3:8)
This FIR filter has L (= 2M÷1) coefficients b
/
l
, l = 0, 1, . . . , L ÷1. The impulse
response is symmetric about b
/
M
due to the fact that h(÷n) = h(n) given in (5.3.4). The
duration of the impulse response is 2MT where T is the sampling period.
From (5.3.6) and (5.3.8), we can show that
B
/
(z) = z
÷M
H
/
(z) (5:3:9)
and
B
/
(!) = e
÷j!M
H
/
(!): (5:3:10)
DESIGN OF FIR FILTERS 203
Since [e
÷j!M
[ = 1, we have
[B
/
(!)[ = [H
/
(!)[: (5:3:11)
This causal filter has the same magnitude response as that of the non-causal filter. If h(n)
is real, then H
/
(!) is a real function of ! (see exercise problem). As discussed in Section
5.2.3, if H
/
(!) _ 0, then the phase of B
/
(!) is equal to ÷M!. If H
/
(!) < 0, then the
phase of B
/
(!) is equal to p ÷M!. Therefore the phase of B
/
(!) is a linear function of !
and thus the transfer function B
/
(z) has a constant group delay.
Example 5.5: The ideal lowpass filter of Figure 5.1(a) has frequency response
H(!) =
1, [![ _ !
c
0, otherwise.

(5:3:12)
The corresponding impulse response can be computed using (5.3.2) as
h(n) =
1
2p

p
÷p
H(!)e
j!n
d! =
1
2p

!
c
÷!
c
e
j!n
d!
=
1
2p
e
j!n
jn
¸
!
c
÷!
c
=
1
2p
e
j!
c
n
÷e
÷j!
c
n
jn
¸
=
sin(!
c
n)
pn
=
!
c
p
sinc
!
c
n
p

,
(5:3:13a)
is referred to as sinc function, where a commonly used precise form for the sinc
function is defined as
sinc(x) =
sin(px)
px
: (5:3:13b)
Taking the limit as n ÷0, we have
h(0) = !
c
=p: (5:3:14)
By setting all impulse response coefficients outside the range ÷M _ n _ M
to zero, we obtain an FIR filter with the symmetry property h
/
(n) = h
/
(÷n),
n = 0, 1, . . . , M. For M = 7, we have ¦

2

=6p, 1=2p,

2

=2p, 1/4,

2

=2p, 1=2p,

2

=6p¦. By shifting M units to the right, we obtain a causal FIR filter of finite-
length L = 2M÷1 with coefficients
b
/
l
=
!
c
p
sinc
!
c
(l ÷M)
p
¸
, 0 _ l _ L ÷1
0, otherwise.

(5:3:15)
Example 5.6: Design a lowpass FIR filter with the frequency response
H( f ) =
1, 0 _ f _ 1 kHz
0, 1 kHz < f _ 4 kHz

204 DESIGN AND IMPLEMENTATION OF FIR FILTERS
when the sampling rate is 8 kHz. The duration of the impulse response is limited to
2.5 msec.
Since 2MT = 0:0025 seconds and T = 0:000125 seconds, we obtain M = 10:
Thus the actual filter has 21 coefficients. From Table 3.1, 1 kHz corresponds to
!
c
= 0:25p. From (5.3.13), we have
h(n) = 0:25 sinc
0:25pn
p

, n = 0, 1, . . . , 10:
Since h(÷n) = h(n), for n = 0, 1, . . . , 10, we can obtain, b
/
l
= h(l ÷10),
l = 0, 1, . . . , 20. The transfer function of the designed causal filter is
B
/
(z) =
¸
20
l=0
b
/
l
z
÷1
:
Example 5.7: Design a lowpass filter of cut-off frequency !
c
= 0:4p with filter
length L = 41 and L = 61.
When L = 41, M = (L ÷1)=2 = 20. From (5.3.15), the designed impulse
response is given by
b
/
l
= 0:4 sinc
0:4p(l ÷20)
p

, l = 0, 1, . . . , 40:
When L = 61, M = (L ÷1)=2 = 30. The impulse response becomes
b
/
l
= 0:4 sinc
0:4p(l ÷30)
p

, l = 0, 1, . . . , 60:
The magnitude responses are computed and plotted in Figure 5.13 using the
MATLAB script exam5_7.m given in the software package.
5.3.3 Gibbs Phenomenon
As shown in Figure 5.13, the causal FIR filter obtained by simply truncating the
impulse response coefficients of the desired filter exhibits an oscillatory behavior
(or ripples) in its magnitude response. As the length of the filter is increased, the number
of ripples in both passband and stopband increases, and the width of the ripples
decrease. The ripple becomes narrower, but its height remains almost constant. The
largest ripple occurs near the transition discontinuity and their amplitude is independent
of L. This undesired effect is called the Gibbs phenomenon. This is an unavoidable
consequence of having an abrupt discontinuity (or truncation) of impulse response in
time domain.
The truncation operation described in (5.3.5) can be considered as multiplication
of the infinite-length sequence {h(n)} by the rectangular sequence {w(n)}. That is,
DESIGN OF FIR FILTERS 205
1
0.8
0.6
0.4
0.2
1
−3 −2 −1 0
M
a
g
n
i
t
u
d
e
Frequency
(a)
(b)
1 2 3
1
0.8
0.6
0.4
0.2
1
−3
M
a
g
n
i
t
u
d
e
−2 −1 0
Frequency
1 2 3
Figure 5.13 Magnitude responses of lowpass filters designed by Fourier series method:
(a) L = 41, and (b) L = 61
h
/
(n) = h(n)w(n), ÷·_ n _ ·, (5:3:16)
where the rectangular window w(n) is defined as
w(n) =
1, ÷M _ n _ M
0, otherwise.

(5:3:17)
The discrete-time Fourier transform (DTFT) of h
/
(n) defined in (5.3.16) can be
expressed as
H
/
(!) = H(!) + W(!) =
1
2p

p
÷p
H(j)W(! ÷j) dj, (5:3:18)
where W(!) is the DTFT of w(n) defined in (5.3.17). Thus the designed filter H
/
(!) will
be a smeared version of the desired filter H(!).
206 DESIGN AND IMPLEMENTATION OF FIR FILTERS
Equation (5.3.18) shows that H
/
(!) is obtained by the convolution of the desired
frequency response H(!) with the rectangular window's frequency response W(!). If
W(! ÷j) = 2pd(! ÷j), (5:3:19)
we have the desired result H
/
(!) = H(!). Equation (5.3.19) implies that if W(!) is a
very narrow pulse centered at ! = 0 such as a delta function W(!) = 2pd(!), then
H
/
(!) will approximate H(!) very closely. From Table 4.1, this condition requires the
optimum window
w(n) = 1, [n[ < ·, (5:3:20)
which has infinite length.
In practice, the length of the window should be as small as possible in order to reduce
the computational complexity of the FIR filter. Therefore we have to use sub-optimum
windows that have the following properties:
1. They are even functions about n = 0.
2. They are zero in the range [n[ > M.
3. Their frequency responses W(!) have a narrow mainlobe and small sidelobes as
suggested by (5.3.19).
The oscillatory behavior of a truncated Fourier series representation of FIR filter,
observed in Figure 5.13, can be explained by the frequency response of rectangular
window defined in (5.3.17). It can be expressed as
W(!) =
¸
M
n=÷M
e
÷j!n
(5:3:21a)
=
sin[(2M÷1)!=2[
sin(!=2)
: (5:3:21b)
A plot of W(!) is illustrated in Figure 5.14 for M = 8 and 20. The MATLAB script
fig5_14.m that generated this figure is available in the software package. The fre-
quency response W(!) has a mainlobe centered at ! = 0. All the other ripples in the
frequency response are called the sidelobes. The magnitude response [W(!)[ has the first
zero at (2M÷1)!=2 = p. That is, ! = 2p=(2M÷1). Therefore the width of the main-
lobe is 4p=(2M÷1). From (5.3.21a), it is easy to show that the magnitude of mainlobe
is [W(0)[ = 2M÷1. The first sidelobe is approximately at frequency !
1
= 3p=(2M÷1)
with magnitude [W(!
1
)[ ~ 2(2M÷1)=3p for M 1. The ratio of the mainlobe mag-
nitude to the first sidelobe magnitude is
W(0)
W(!
1
)

~
3p
2
= 13:5 dB: (5:3:22)
DESIGN OF FIR FILTERS 207
40
20
0
M
a
g
n
i
t
u
d
e

(
d
B
)
−20
−40
−3 −2 −1 0 1 2 3
40
20
0
M
a
g
n
i
t
u
d
e

(
d
B
)
−20
−40
−3 −2 −1 0
Frequency
1 2 3
Figure 5.14 Frequency response of the rectangular window for M = 8 (top) and 20 (bottom)
As ! increases to the Nyquist frequency, p, the denominator grows larger. This attenu-
ates the higher-frequency numerator terms, resulting in the damped sinusoidal function
shown in Figure 5.14.
As M increases, the width of the mainlobe decreases as desired. However, the area
under each lobe remains constant, while the width of each lobe decreases with an
increase in M. This implies that with increasing M, ripples in H
/
(!) around the point
of discontinuity occur more closely but with no decrease in amplitude.
The rectangular window has an abrupt transition to 0 outside the range
÷M _ n _ M, which causes the Gibbs phenomenon in the magnitude response of the
windowed filter's impulse response. The Gibbs phenomenon can be reduced by either
using a window that tapers smoothly to 0 at each end, or by providing a smooth
transition from the passband to the stopband. A tapered window causes the height of
the sidelobes to diminish and increases in the mainlobe width, resulting in a wider
transition at the discontinuity. This phenomenon is often referred to as leakage or
smearing.
5.3.4 Window Functions
A large number of tapered windows have been developed and optimized for different
applications. In this section, we restrict our discussion to four commonly used windows
of length L = 2M÷1. That is, w(n), n = 0, 1, . . . , L ÷1 and is symmetric about its
middle, n = M. Two parameters that predict the performance of the window in FIR
filter design are its mainlobe width and the relative sidelobe level. To ensure a fast
transition from the passband to the stopband, the window should have a small mainlobe
width. On the other hand, to reduce the passband and stopband ripples, the area under
208 DESIGN AND IMPLEMENTATION OF FIR FILTERS
the sidelobes should be small. Unfortunately, there is a trade-off between these two
requirements.
The Hann (Hanning) window function is one period of the raised cosine function
defined as
w(n) = 0:5 1 ÷cos
2pn
L ÷1
¸
, n = 0, 1, . . . , L ÷1: (5:3:23)
Note that the Hanning window has an actual length of L ÷2 since the two end values
given by (5.3.23) are zero. The window coefficients can be generated by the MATLAB
built-in function
w = hanning(L);
which returns the L-point Hanning window function in array w. Note that the
MATLAB window functions generate coefficients w(n), n = 1, . . . , L, and is shown in
Figure 5.15 (top). The magnitude response of the Hanning window is shown in the
bottom of Figure 5.15. The MATLAB script han.m is included in the software package.
For a large L, the peak-to-sidelobe ratio is approximately 31 dB, an improvement of
17.5 dB over the rectangular window. However, since the width of the transition band
corresponds roughly to the mainlobe width, it is more than twice that resulting from the
rectangular window shown in Figure 5.14.
The Hamming window function is defined as
w(n) = 0:54 ÷0:46 cos
2pn
L ÷1

, n = 0, 1, . . . , L ÷1, (5:3:24)
1
0.8
0.6
0.4
0.2
0
A
m
p
l
i
t
u
d
e
0 5 10 15 20 25 30 35 40
40
20
0
−20
−40
−60
−80
M
a
g
n
i
t
u
d
e

(
d
B
)
−3 −2 −1 0
Frequency
Time Index
1 2 3
Figure 5.15 Hanning window function (top) and its magnitude response (bottom), L = 41
DESIGN OF FIR FILTERS 209
which also corresponds to a raised cosine, but with different weights for the constant
and cosine terms. The Hamming function does not taper the end values to 0, but rather
to 0.08. MATLAB provides the Hamming window function as
w = hamming(L);
This window function and its magnitude response are shown in Figure 5.16, and the
MATLAB script ham.m is given in the software package.
The mainlobe width is about the same as for the Hanning window, but has
an additional 10 dB of stopband attenuation (41 dB). In designing a lowpass filter,
the Hamming window provides low ripple over the passband and good stopband
attenuation is usually more appropriate for FIR filter design than the Hanning window.
Example 5.8: Design a lowpass filter of cut-off frequency !
c
= 0:4p and order
L = 61 using the Hamming window. Using MATLAB script (exam5_8.m in the
software package) similar to the one used in Example 5.7, we plot the magnitude
response in Figure 5.17. Compared with Figure 5.13(b), we observe that the
ripples produced by rectangular window design are virtually eliminated from the
Hamming window design. The trade-off for eliminating the ripples is loss of
resolution, which is shown by increasing transition width.
The Blackman window function is defined as
w(n) = 0:42 ÷0:5 cos
2pn
L ÷1

÷0:08 cos
4pn
L ÷1

, (5:3:25)
n = 0, 1, . . . , L ÷1:
1
0.8
0.6
0.4
0.2
0
A
m
p
l
i
t
u
d
e
0 5 10 15 20 25 30 35 40
40
20
0
−20
−40
−60
−80
M
a
g
n
i
t
u
d
e

(
d
B
)
−3 −2 −1 0
Frequency
Time Index
1 2 3
Figure 5.16 Hamming window function (top) and its magnitude response (bottom), L = 41
210 DESIGN AND IMPLEMENTATION OF FIR FILTERS
1
0.8
0.6
0.4
0.2
0
M
a
g
n
i
t
u
d
e

(
d
B
)
−3 −2 −1 0
Frequency
1 2 3
Figure 5.17 Magnitude response of lowpass filter using Hamming window, L = 61
1
0.8
0.6
0.4
0.2
0
A
m
p
l
i
t
u
d
e
0 5 10 15 20 25 30 35 40
40
20
0
−20
−40
−60
−80
M
a
g
n
i
t
u
d
e

(
d
B
)
−3 −2 −1 0
Frequency
Time Index
1 2 3
Figure 5.18 Blackman window function (top) and its magnitude response (bottom), L = 41
This function is also supported by the MATLAB function
w = blackman(L);
This window and its magnitude response are shown in Figure 5.18 using the MATLAB
script bmw.m given in the software package.
The addition of the second cosine term in (5.3.25) has the effect of increasing the
width of the mainlobe (50 percent), but at the same time improving the peak-to-sidelobe
DESIGN OF FIR FILTERS 211
ratio to about 57 dB. The Blackman window provides 74 dB of stopband attenuation,
but with a transition width six times that of the rectangular window.
The Kaiser window function is defined as
w(n) =
I
0
b

1 ÷(n ÷M)
2
=M
2

¸
I
0
(b)
, n = 0, 1, . . . , L ÷1, (5:3:26a)
where b is an adjustable (shape) parameter and
I
0
(b) =
¸
·
k=0
(b=2)
k
k!
¸ ¸
2
(5:3:26b)
is the zero-order modified Bessel function of the first kind. In practice, it is sufficient to
keep only the first 25 terms in the summation of (5.3.26b). Because I
0
(0) = 1, the Kaiser
window has the value 1=I
0
(b) at the end points n = 0 and n = L ÷1, and is symmetric
about its middle n = M. This is a useful and very flexible family of window functions.
MATLAB provides Kaiser window function as
kaiser(L, beta);
The window function and its magnitude response are shown in Figure 5.19 for L = 41
and b = 8 using the MATLAB script ksw.m given in the software package. The Kaiser
window is nearly optimum in the sense of having the most energy in the mainlobe for a
given peak sidelobe level. Providing a large mainlobe width for the given stopband
attenuation implies the sharpness transition width. This window can provide different
transition widths for the same L by choosing the parameter b to determine the trade-off
between the mainlobe width and the peak sidelobe level.
As shown in (5.3.26), the Kaiser window is more complicated to generate, but the
windowfunction coefficients are generated only once during filter design. Since a window
is applied to each filter coefficient when designing a filter, windowing will not affect the
run-time complexity of the designed FIR filter.
Although d
p
and d
s
can be specified independently as given in (5.1.8) and (5.1.9), FIR
filters designed by all windows will have equal passband and stopband ripples. There-
fore we must design the filter based on the smaller of the two ripples expressed as
d = min(d
p
, d
s
): (5:3:27)
The designed filter will have passband and stopband ripples equal to d. The value of d
can be expressed in dB scale as
A = ÷20 log
10
d dB: (5:3:28)
In practice, the design is usually based on the stopband ripple, i.e., d = d
s
. This is because
any reasonably good choices for the passband and stopband attenuation (such as
A
p
= 0:1 dB and A
s
= 60 dB) will result in d
s
< d
p
.
212 DESIGN AND IMPLEMENTATION OF FIR FILTERS
1
0.8
0.6
0.4
0.2
0
A
m
p
l
i
t
u
d
e
0 5 10 15 20 25 30 35 40
40
20
0
−20
−40
−60
−80
M
a
g
n
i
t
u
d
e

(
d
B
)
−3 −2 −1 0
Frequency
Time Index
1 2 3
Figure 5.19 Kaiser window function (top) and its magnitude response (bottom), L = 41 and
b = 8
The main limitation of Hanning, Hamming, and Blackman windows is that they
produce a fixed value of d. They limit the achievable passband and stopband attenu-
ation to only certain specific values. However, the Kaiser window does not suffer from
the above limitation because it depends on two parameters, L and b, to achieve any
desired value of ripple d or attenuation A. For most practical applications, A _ 50 dB.
The Kaiser window parameters are determined in terms of the filter specifications d and
transition width Df as follows [5]:
b = 0:1102(A ÷8:7) (5:3:29)
and
L =
(A ÷7:59)f
s
14:36Df
÷1, (5:3:30)
where Df = f
stop
÷f
pass
.
Example 5.9: Design a lowpass filter using the Kaiser window with the follow-
ing specifications: f
s
= 10 kHz, f
pass
= 2 kHz, f
stop
= 2:5 kHz, A
p
= 0:1 dB, and
A
s
= 80 dB.
From (5.1.8), d
p
=
10
0:1=20
÷1
10
0:1=20
÷1
= 0:0058. From (5.1.9), d
s
= 10
÷80=20
= 0:0001.
Thus we choose d = d
s
= 0:0001 from (5.3.27) and A = ÷20 log
10
d = 80 = A
s
from (5.3.28). The shaping parameter b is computed as
DESIGN OF FIR FILTERS 213
b = 0:1102(80 ÷8:7) = 7:875:
The required window length L is computed as
L =
(80 ÷7:59)10
14:36(2:5 ÷2)
÷1 ~ 101:85:
Thus we choose the filter order L = 103.
The procedures of designing FIR filters using windows are summarized as follows:
1. Determine the window type that will satisfy the stopband attenuation requirements.
2. Determine the window size L based on the given transition width.
3. Calculate the window coefficients w(n), n = 0, 1, . . . , L ÷1.
4. Generate the ideal impulse response h(n) using (5.3.3) for the desired filter.
5. Truncate the ideal impulse response of infinite length using (5.3.5) to obtain
h
/
(n), ÷M _ n _ M.
6. Make the filter causal by shifting the result M units to the right using (5.3.7) to
obtain b
/
l
, l = 0, 1, . . . , L ÷1.
7. Multiply the window coefficients obtained in step 3 and the impulse response
coefficients obtained in step 6 sample-by-sample. That is,
b
l
= b
/
l
w(l), l = 0, 1, . . . , L ÷1: (5:3:31)
Applying a window to an FIR filter's impulse response has the effect of smoothing the
resulting filter's magnitude response. A symmetric window will preserve a symmetric
FIR filter's linear-phase response.
The advantage of the Fourier series method with windowing is its simplicity. It does
not require sophisticated mathematics and can be carried out in a straightforward
manner. However, there is no simple rule for choosing M so that the resulting filter
will meet exactly the specified cut-off frequencies. This is due to the lack of an exact
relation between M and its effect on leakage.
5.3.5 Frequency Sampling Method
The frequency-sampling method is based on sampling a desired amplitude spectrum and
obtaining filter coefficients. In this approach, the desired frequency response H(!) is
first uniformly sampled at L equally spaced points !
k
=
2pk
L
, k = 0, 1, . . . , L ÷1. The
frequency-sampling technique is particularly useful when only a small percentage of the
214 DESIGN AND IMPLEMENTATION OF FIR FILTERS
frequency samples are non-zero, or when several bandpass functions are desired
simultaneously. A unique attraction of the frequency-sampling method is that it also
allows recursive implementation of FIR filters, leading to computationally efficient
filters. However, the disadvantage of this method is that the actual magnitude response
of the filter will match the desired filter only at the points that were samples.
For a given frequency response H(!), we take L samples at frequencies of kf
s
=L,
k = 0, 1, . . . , L ÷1 to obtain H(k), k = 0, 1, . . . L ÷1. The filter coefficients b
l
can be
obtained as the inverse DFT of these frequency samples. That is,
b
l
=
1
L
¸
L÷1
k=0
H(k)e
j(2p=L)lk
, l = 0, 1, . . . , L ÷1: (5:3:32)
The resulting filter will have a frequency response that is exactly the same as the original
response at the sampling instants. However, the response may be significantly different
between the samples. To obtain a good approximation to the desired frequency
response, we have to use a sufficiently large number of frequency samples. That is, we
have to use a large L.
Let ¦B
k
¦ be the DFT of ¦b
l
¦ so that
B
k
=
¸
L÷1
l=0
b
l
e
÷j(2p=L)lk
, k = 0, 1, . . . , L ÷1 (5:3:33)
and
b
l
=
1
L
¸
L÷1
k=0
B
k
e
j(2p=L)lk
, l = 0, 1, . . . , L ÷1: (5:3:34)
Using the geometric series
¸
L÷1
l=0
x
l
=
1 ÷x
L
1 ÷x
(see Appendix A.2), the desired filter's
transfer function can be obtained as
H(z) =
¸
L÷1
l=0
b
l
z
÷l
=
¸
L÷1
l=0
1
L
¸
L÷1
k=0
B
k
e
j(2plk=L)

z
÷l
=
1
L
(1 ÷z
÷L
)
¸
L÷1
k=0
B
k
1 ÷e
j(2pk=L)
z
÷1
: (5:3:35)
This equation changes the transfer function into a recursive form, and H(z) can be
viewed as a cascade of two filters: a comb filter, (1 ÷z
÷L
)=L as discussed in Section
5.2.2, and a sum of L first-order all-pole filters.
The problem is now to relate ¦B
k
¦ to the desired sample set {H(k)} used in (5.3.32). In
general, the frequency samples H(k) are complex. Thus a direct implementation of
(5.3.35) would require complex arithmetic. To avoid this complication, we use the
symmetry inherent in the frequency response of any FIR filter with real impulse
response h(n). Suppose that the desired amplitude response [H(!)[ is sampled such
DESIGN OF FIR FILTERS 215
that B
k
and H(k) are equal in amplitude. For a linear-phase filter (assume positive
symmetry), we have
H
k
= [H(k)[ = [H(L ÷k)[ = [B
k
[, k _
L
2
: (5:3:36)
The remaining problem is to adjust the relative phases of the B
k
so that H
k
will
provide a smooth approximation to [H(!)[ between the samples. The phases of two
adjacent contributions to [H(!)[ are in phase except between the sample points, which
are 180 degrees out of phase [10]. Thus the two adjacent terms should be subtracted to
provide a smooth reconstruction between sample points. Therefore B
k
should be equal
to H
k
with alternating sign. That is,
B
k
= (÷1)
k
H
k
, k _
L
2
: (5:3:37)
This is valid for L being odd or even, although it is convenient to assume that B
0
is zero
and B
L=2
is 0 if L is even.
With these assumptions, the transfer function given in (5.3.35) can be further
expressed as
H(z) =
1
L
(1 ÷z
÷L
)
¸
k_L=2
(÷1)
k
H
k
1
1 ÷e
j(2pk=L)
z
÷1
÷
1
1 ÷e
j[2p(L÷k)=L[
z
÷1
¸
=
2
L
(1 ÷z
÷L
)
¸
k_L=2
(÷1)
k
H
k
1 ÷ cos(2pk=L)z
÷1
1 ÷2 cos(2pk=L)z
÷1
÷z
÷2
: (5:3:38)
This equation shows that H(z) has poles at e
±j(2pk=L)
on the unit circle in the z-plane. The
comb filter 1 ÷z
÷L
provides L zeros at z
k
= e
j(2pk=L)
, k = 0, 1, . . . , L ÷1, equally
spaced around the unit circle. Each non-zero sample H
k
brings to the digital filter a
conjugate pair of poles, which cancels a corresponding pair of zeros at e
±j(2pk=L)
on the
unit circle. Therefore the filter defined in (5.3.38) is recursive, but does not have poles in
an essential sense. Thus it still has a finite impulse response.
Although the pole±zero cancellation is exact in theory, it will not be in practice due to
finite wordlength effects in the digital implementation. The cos(2pk=L) terms in (5.3.38)
will in general be accurate only if more bits are used. An effective solution to this
problem is to move the poles and zeros slightly inside the unit circle by moving them
into radius r, where r < 1.
Therefore (5.3.38) is modified to
H(z) =
2
L
(1 ÷r
L
z
÷L
)
¸
k_L=2
(÷1)
k
H
k
1 ÷r cos(2pk=L)z
÷1
1 ÷2r cos(2pk=L)z
÷1
÷r
2
z
÷2
: (5:3:39)
The frequency-sampling filter defined in (5.3.39) can be realized by means of a comb
filter
216 DESIGN AND IMPLEMENTATION OF FIR FILTERS
C(z) =
2
L
(1 ÷r
L
z
÷L
) (5:3:40)
and a bank of resonators
R
k
(z) =
1 ÷r cos(2pk=L)z
÷1
1 ÷2r cos(2pk=L)z
÷1
÷r
2
z
÷2
, 0 _ k _ L=2, (5:3:41)
as illustrated in Figure 5.20. They effectively act as narrowband filters, each passing
only those frequencies centered at and close to resonant frequencies 2pk=L and exclud-
ing others outside these bands. Banks of these filters weighted by frequency samples H
k
can then be used to synthesize a desired frequency response.
The difference equation representing the comb filter C(z) can be written in terms of
variables in Figure 5.20 as
u(n) =
2
L
[x(n) ÷r
L
x(n ÷L)[: (5:3:42)
The block diagram of this modified comb filter is illustrated in Figure 5.21. An effective
technique to implement this comb filter is to use the circular buffer, which is available in
most modern DSP processors such as the TMS320C55x. The comb filter output u(n) is
common to all the resonators R
k
(z) connected in parallel.
The resonator with u(n) as common input can be computed as
f
k
(n) = u(n) ÷r cos
2pk
L

u(n ÷1) ÷2r cos
2pk
L

f
k
(n ÷1) ÷r
2
f
k
(n ÷2)
= u(n) ÷r cos
2pk
L

[2f
k
(n ÷1) ÷u(n ÷1)[ ÷r
2
f
k
(n ÷2), 0 _ k _ L=2 (5:3:43)
Comb
filter
y(n) x(n)
C(z)
u(n)
From other channel
From other channel
R
0
(z)
R
k
(z)
f
k
(n)
H
k
R
L/2
(z)
Σ
Figure 5.20 Block diagram of frequency sampling filter structure, channel k
x(n)
2/L
u(n)
+

r
L
z
−L
Σ
Figure 5.21 Detailed block diagram of comb filter
DESIGN OF FIR FILTERS 217

u(n)
z
−1
z
−1
z
−1
f
k
(n)
f
k
(n−1)
f
k
(n−2)






L
rcos
2pk
2
r
2
Σ Σ

Figure 5.22 Detailed block diagram of resonator, R
k
(z)
where f
k
(n) is the output from the kth resonator. The detailed flow diagram of the
resonator is illustrated in Figure 5.22. Note that only one coefficient r cos
2pk
L

is
needed for each resonator. This significantly reduces the memory requirements when
compared to other second-order IIR bandpass filters.
Finally, the output of each resonator f
k
(n) is weighted by H
k
and combined into the
overall output
y(n) =
¸
k_L=2
(÷1)
k
H
k
f
k
(n): (5:3:44)
Example 5.10: Design a bandpass filter of sampling rate 10 kHz, passband
1±1.25 kHz, power gain at ends of passband is 0.5 (÷3 dB), and duration of
impulse response is 20 ms.
Since the impulse response duration is 20 ms,
L = 0:02=T = 200:
The frequency samples must be 10 000/200 = 50 Hz apart. The passband is
sampled at frequencies 1, 1.05, 1.1, 1.15, 1.2, and 1.25 kHz that correspond to
k = 20, 21, 22, 23, 24, and 25.
Therefore the frequency-sampling filter consists of six channels (resonators) as
shown in Figure 5.20 with
H
20
= H
25
=

0:5

= 0:707 (÷3dB) and H
21
= H
22
= H
23
= H
24
= 1:
The designed frequency-sampling filter is implemented in C (fsf.c in the soft-
ware package) using the circular buffer. In the program, we assume the input data
file is using binary floating-point format, and save the output using the same
format.
218 DESIGN AND IMPLEMENTATION OF FIR FILTERS
5.4 Design of FIR Filters Using MATLAB
The filter design methods described in Section 5.3 can be easily realized using a
computer. A number of filter design algorithms are based on iterative optimization
techniques that are used to minimize the error between the desired and actual frequency
responses. The most widely used algorithm is the Parks±McClellan algorithm for design-
ing the optimum linear-phase FIR filter. The algorithm spreads out the error within a
frequency band, which produces ripples of equal magnitude. The user can specify the
relative importance of each band. For example, there would be less ripples in the pass-
band than in the stopband. A variety of software packages are commercially available
that have made digital filter design rather simple on a computer. In this section, we
consider only the MATLAB application for designing FIR filters.
As discussed in Section 5.1.3, the filter specifications !
p
, !
s
, d
p
, and d
s
are given. The
filter order L can be estimated using the simple formula
L = 1 ÷
÷20 log
10

d
p
d
s

÷13
14:6Df
, (5:4:1)
where
Df =
!
s
÷!
p
2p
: (5:4:2)
A highly efficient procedure, the Remez algorithm, is developed to design the opti-
mum linear-phase FIR filters based on the Parks±McClellan algorithm. The algorithm
uses the Remez exchange and Chebyshev approximation theory to design a filter with
an optimum fit between the desired and actual frequency responses. This algorithm is
implemented as an M-file function remez that is available in the Signal Processing
Toolbox of MATLAB. There are various versions of this function:
b = remez(N, f, m);
b = remez(N, f, m, w);
b = remez(N, f, m, `ftype');
b = remez(N, f, m, w, `ftype');
The function returns row vector b containing the N ÷1 coefficients of the FIR filter of
order L = (N ÷1). The vector f specifies bandedge frequencies in the range between 0
and 1, where 1 corresponds to the Nyquist frequency f
N
= f
s
=2. The frequencies must be
in an increasing order with the first element being 0 and the last element being 1. The
desired values of the FIR filter magnitude response at the specified bandedge frequen-
cies in f are given in the vector m, with the elements given in equal-valued pairs. The
vector f and m must be the same length with the length being an even number.
Example 5.11: Design a linear-phase FIR bandpass filter of order 18 with a
passband from 0.4 to 0.6. Plot the desired and actual frequency responses using
the following MATLAB script (exam5_11.m in the software package):
f = [0 0.3 0.4 0.6 0.7 1];
m = [0 0 1 1 0 0];
DESIGN OF FIR FILTERS USING MATLAB 219
b = remez(17, f, m);
[h, omega]= freqz(b, 1, 512);
plot(f, m, omega/pi, abs(h));
The graph is shown in Figure 5.23.
The desired magnitude response in the passband and the stopband can be weighted by
an additional vector w. The length of w is half of the length of f and m. As shown in
Figure 5.7, there are four types of linear-phase FIR filters. Types III (L even) and IV (L
odd) are used for specialized filter designs: the Hilbert transformer and differentiator.
To design these two types of FIR filters, the arguments `hilbert' and `differen-
tiator' are used for `ftype' in the last two versions of remez.
Similar to remez, MATLAB also provides firls function to design linear-phase
FIR filters which minimize the weighted, integrated squared error between the ideal
filter and the actual filter's magnitude response over a set of desired frequency bands.
The synopsis of function firls is identical to the function remez.
Two additional functions available in the MATLAB Signal Processing Toolbox,
fir1 and fir2, can be used in the design of FIR filters using windowed Fourier series
method. The function fir1 designs windowed linear-phase lowpass, highpass, band-
pass, and bandstop FIR filters with the following forms:
b = fir1(N, Wn);
b = fir1(N, Wn, `filtertype');
b = fir1(N, Wn, window);
b = fir1(N, Wn, `filtertype', window);
The basic form, b = fir1(N, Wn), generates the length L = N ÷1 vector b containing
the coefficients of a lowpass filter with a normalized cut-off frequency Wn between 0 and
1.4
1.2
1
0.8
0.6
0.4
0.2
0
0 0.1 0.2 0.3 0.4
M
a
g
n
i
t
u
d
e
0.5
Normalized frequency
0.6 0.7 0.8 0.9 1
Actual filter
Ideal filter
Figure 5.23 Magnitude responses of the desired and actual FIR filters
220 DESIGN AND IMPLEMENTATION OF FIR FILTERS
1 using the Hamming window. If Wn is a two-element vector, Wn = [w1 w2], fir1
generates a bandpass filter with passband w1 _ ! _ w2. The argument `filtertype'
specifies a filter type, where `high' for a highpass filter with cut-off frequency Wn, and
`stop' for a bandstop filter with cut-off frequency Wn = [w1 w2]. The vector window
must have L elements and is generated by
window = hamming(L);
window = hanning(L);
window = blackman(L);
window = kaiser(L, beta);
If no window is specified, the Hamming window will be used as the default for filter
design.
The function fir2 is used to design an FIR filter with arbitrarily shaped magnitude
response. The various forms of this function are:
b = fir2(N, f, m);
b = fir2(N, f, m, window);
b = fir2(N, f, m, npt);
b = fir2(N, f, m, npt, window);
b = fir2(N, f, m, npt, lap);
b = fir2(N, f, m, npt, lap, window);
The basic form, b = fir2(N, f, m), is used to design an FIR filter of order L = N ÷1.
The vector m defines the magnitude response sampled at the specified frequency points
given by the vector f. The argument npt specifies the number of points for the grid
onto which fir2 interpolates the frequency response. The argument lap specifies the
size of the region which fir2 inserts around duplicated frequency points. The details
are given in the Signal Processing Toolbox manual.
5.5 Implementation Considerations
Discrete-time FIR filters designed in the previous section can be implemented in the
following forms: hardware, firmware, and software. Digital filters are realized by digital
computers or DSP hardware uses quantized coefficients to process quantized signals. In
this section, we discuss the software implementation of digital FIR filters using
MATLAB and C to illustrate the main issues. We will consider finite wordlength effects
in this section. The DSP chip implementation using the TMS320C55x will be presented
in the next section.
5.5.1 Software Implementations
The software implementation of digital filtering algorithms on a PC is often carried out
first to verify whether or not the chosen algorithm meets the goals of the application.
MATLAB and C implementation may be adequate if the application does not require
real-time signal processing. For simulation purposes, it is convenient to use a powerful
software package such as MATLAB for signal processing.
IMPLEMENTATION CONSIDERATIONS 221
MATLAB provides the built-in function filter for FIR and IIR filtering. The basic
form of this function is
y = filter(b, a, x)
For FIR filtering, a =1 and filter coefficients b
l
are contained in the vector b. The input
vector is x while the output vector generated by the filter is y.
Example 5.12: The following C function fir.c implements the linear convolution
(FIR filtering, inner product, or dot product) operation given in (5.2.1). The
arrays x and h are declared to proper dimension in the main programfirfltr.c
given in Appendix C.
/**************************************************************
* FIR ÷ This function performs FIR filtering (linear convolution)
* ntap-1
* y(n)= sum hi*x(n÷i)
* i=0
**************************************************************/
float fir(float *x, float *h, int ntap)
{
float yn = 0.0; /* Output of FIR filter */
int i; /* Loop index */
for(i = 0; i > ntap; i++)
{
yn ÷= h [i]*x [i]; /* Convolution of x(n) with h(n) */
}
return(yn); /* Return y(n) to main function */
}
The signal buffer x is refreshed in every sampling period as shown in Figure
5.11, and is implemented in the following C function called shift.c. At each
iteration, the oldest sample x(n ÷L ÷1) discarded and other signals are shifted
one location to the right. The most recent sample x(n) is inserted to memory
location x [0].
/**************************************************************
* SHIFT ÷ This function updates signal vector of order ntap
* data stored as [x(n) x(n÷1) ... x(n÷ntap+1)]
**************************************************************/
void shift(float *x, int ntap, float in)
{
int i; /* Loop index */
for(i = ntap÷1; i > 0; i÷÷)
{
x [i]= x [i÷1]; /* Shift old data x(n÷i) */
}
x [0]= in; /* Insert new data x(n) */
return;
}
222 DESIGN AND IMPLEMENTATION OF FIR FILTERS
The FIR filtering defined in (5.2.1) can be implemented using DSP chips or special
purpose ASIC devices. Modern programmable DSP chips, such as the TMS320C55x,
have architecture that is optimized for the repetitive nature of multiplications and
accumulations. They are also optimized for performing the memory moves required
in updating the contents of the signal buffer, or realizing the circular buffers. The
implementation of FIR filters using the TMS320C55x will be discussed in Section 5.6.
5.5.2 Quantization Effects in FIR Filters
Consider an FIR filter transfer function given in (5.2.2). The filter coefficients, b
l
, are
determined by a filter design package such as MATLAB for given specifications. These
coefficients are usually represented by double-precision floating-point numbers and
have to be quantized using a finite number of bits for a given fixed-point processor
such as 16-bit for the TMS320C55x. The filter coefficients are only quantized once in
the design process, and those values remain constant in the filter implementation. We
must check the quantized design. If it no longer meets the given specifications, we can
optimize, redesign, restructure, and/or use more bits to satisfy the specifications. It is
especially important to consider quantization effects when designing filters for imple-
mentation using fixed-point arithmetic.
Let b
/
l
denote the quantized values corresponding to b
l
. As discussed in Chapter 3, the
nonlinear quantization can be modeled as a linear operation expressed as
b
/
l
= Q[b
l
[ = b
l
÷e(l), (5:5:1)
where e(l ) is the quantization error and can be assumed as a uniformly distributed
random noise of zero mean and variance defined in (3.5.6).
Quantization of the filter coefficients results in a new filter transfer function
H
/
(z) =
¸
L÷1
l=0
b
/
l
z
÷l
=
¸
L÷1
l=0
[b
l
÷e(l)[z
÷l
= H(z) ÷E(z), (5:5:2)
where
E(z) =
¸
L÷1
l=0
e(l)z
÷l
(5:5:3)
is the FIR filter representing the error in the transfer function due to coefficient
quantization. The FIR filter with quantized coefficients can be modeled as a parallel
connection of two FIR filters as illustrated in Figure 5.24.
The frequency response of the actual FIR filter with quantized coefficients b
/
l
can be
expressed as
H
/
(!) = H(!) ÷E(!), (5:5:4)
IMPLEMENTATION CONSIDERATIONS 223
y(n)
e(n)
H(z)
E(z)
x(n) yЈ(n)
Figure 5.24 Model of the FIR filter with quantized coefficients
where
E(!) =
¸
L÷1
l=0
e(l)e
÷j!l
(5:5:5)
represents the error in the desired frequency response H(!). The error is bounded by
[E(!)[ =
¸
L÷1
l=0
e(l)e
÷j!l

_
¸
L÷1
l=0
[e(l)[[e
÷j!l
[ _
¸
L÷1
l=0
[e(l)[: (5:5:6)
As shown in (3.5.3),
[e(l)[ _
D
2
= 2
÷B
: (5:5:7)
Thus Equation (5.5.6) becomes
[E(!)[ _ L 2
÷B
: (5:5:8)
This bound is too conservative because it can only be reached if all errors e(l ) are of
the same sign and have the maximum value in the range. A more realistic bound can be
derived assuming e(l ) are statistically independent random variables. The variance of
E(!) can be obtained as
s
2
E
(!) _ 2
÷B÷1

2L ÷1
12

: (5:5:9)
This bound can be used to estimate the wordlength of the FIR coefficients required to
meet the given filter specifications.
As discussed in Section 3.6.3, the most effective technique in preventing overflow is
scaling down the magnitude of signals. The scaling factor used to prevent overflow in
computing the sum of products defined in (5.2.1) is given in (3.6.4) or (3.6.5).
As discussed in Section 5.2.3, most FIR filters are linear-phase and the coefficients are
constrained to satisfy the symmetry condition (5.2.15) or the anti-symmetry property
(5.2.18). Quantizing both sides of (5.2.15) or (5.2.18) has the same quantized value for
224 DESIGN AND IMPLEMENTATION OF FIR FILTERS
each l, which implies that the filter still has linear phase after quantization. Only the
magnitude response of the filter is changed. This constraint greatly reduces the sen-
sitivity of the direct-form FIR filter implementation given in (5.2.1). There is no need to
use the cascade form shown in Figure 5.9 for FIR filters, unlike the IIR filters that
require cascade form. This issue will be discussed in the next chapter.
5.6 Experiments Using the TMS320C55x
FIR filters are widely used in a variety of areas such as audio, video, wireless com-
munications, and medical devices. For many practical applications such as wireless
communications (CDMA/TDMA), streamed video (MPEG/JPEG), and voice over
internet protocol (VoIP), the digital samples are usually grouped in frames with time
duration from a few milliseconds to several hundred milliseconds. It is more efficient
for the C55x to process samples in frames (blocks). The FIR filter programfir.c given
in Section 5.5.1 are designed for processing signals sample-by-sample. It can be
easily modified to handle a block of samples. We call the filter that processes signals
block-by-block a block filter. Example 5.13 is an example of a block-FIR filter written
in C.
Example 5.13: The following C function, block_fir.c, implements an L-tap
FIR filter that processes a block of M input signals at a time.
/****************************************************************
* BLOCK_FIR ÷ This function performs block FIR filter of size M
****************************************************************/
void fir(float *in, int M, float *h, int L, float *out, float *x)
{
float yn; /* Output of FIR filter */
int i,j; /* Loop index */
for(j = 0; j < M; j÷÷)
{
x [0]= in [j]; /* Insert new data x(n) */
/************************************************************
* FIR filtering (linear convolution)
* L÷1
* y(n)= sum hi*x(n÷i)
* i=0
************************************************************/
for(yn = 0.0, i = 0; i < L; i÷÷)
{
yn ÷= h[i]*x[i]; /* Convolution of x(n) with h(n) */
}
out[j]= yn; /* Store y(n) to output buffer */
/************************************************************
* Updates signal buffer, as [x(n) x(n÷1)...x(n÷L+1)]
************************************************************/
EXPERIMENTS USING THE TMS320C55X 225
for(i = L÷1; i > 0; i÷÷)
{
x[i]= x[i÷1]; /* Shift old data x(n÷i) */
}
}
return;
}
Simulation and emulation methods are commonly used in DSP software develop-
ment. They are particularly useful for the study and analysis of DSP algorithms. By
using a software signal generator, we can produce the exact same signals repeatedly
during the debug and analysis processes. Table 5.1 lists the example of sinusoid signal
generator, signal_gen.c that is used to generate the experimental data input5.dat
for experiments in this section.
Table 5.1 List of sinusoid signal generator for the experiments
/*
signal_gen.c ÷ Generate sinewaves as testing data in Q15 format
Prototype: void signal_gen(int *, int)
arg0: ÷ data buffer pointer for output
arg1: ÷ number of samples
*/
#include <math.h>
#define T 0.000125 /* 8000Hz sampling frequency */
#define f1 800 /* 800Hz frequency */
#define f2 1800 /* 1800Hz frequency */
#define f3 3300 /* 3300Hz frequency */
#define PI 3.1415926
#define two_pi_f1_T(2*PI*f1*T) /* 2*pi*f1/Fs */
#define two_pi_f2_T(2*PI*f2*T) /* 2*pi*f2/Fs */
#define two_pi_f3_T(2*PI*f3*T) /* 2*pi*f3/Fs */
#define a1 0.333 /* Magnitude for wave 1 */
#define a2 0.333 /* Magnitude for wave 2 */
#define a3 0.333 /* Magnitude for wave 3 */
static unsigned int n = 0;
void signal_gen(int *x, int N)
{
float temp;
int i;
for(i = 0; i < N; i÷÷)
{
temp = a1*cos((double)two_pi_f1_T*n);
temp ÷= a2*cos((double)two_pi_f2_T*n);
temp ÷= a3*cos((double)two_pi_f3_T*n);
226 DESIGN AND IMPLEMENTATION OF FIR FILTERS
Table 5.1 (continued )
n÷÷;
x[i] = (int)((0x7fff*temp)÷0.5);
}
}
5.6.1 Experiment 5A ± Implementation of Block FIR Filter
The difference equation (5.2.1) and Example 5.13 show that FIR filtering includes two
different processes: (1) It performs a summation of products generated by multiplying
the incoming signals with the filter coefficients. (2) The entire signal buffer is updated to
include a new sample. For an FIR filter of L coefficients, L multiplications, (L ÷1)
additions, and additional data memory move operations are required for the complete
filtering operations. Refreshing the signal buffer in Example 5.13 uses the memory shift
shown in Figure 5.11. To move (L ÷1) samples in the signal buffer to the next memory
location requires additional instruction cycles. These extensive operations make FIR
filtering a computation-intensive task for general-purpose microprocessors.
The TMS320C55x has three important features to support FIR filtering. It has
multiply±accumulate instructions, the circular addressing modes, and the zero-overhead
nested loops. Using multiply±accumulate instructions, the C55x can perform both
multiplication and addition with rounding options in one cycle. That is, the C55x can
complete the computation of one filter tap at each cycle. In Example 5.13, the updating
of the signal buffers by shifting data in memory requires many data-move operations. In
practice, we can use the circular buffers as shown in Figure 5.12. The FIR filtering in
Example 5.13 can be tightly placed into the loops. To reduce the overhead of loop
control, the loop counters in the TMS320C55x are handled using hardware, which can
support three levels of zero-overhead nested loops using BRC0, BRC1, and CSR
registers.
The block FIR filter in Example 5.13 can be implemented with circular buffers using
the following TMS320C55x assembly code:
mov # M÷1,BRC0
mov # L÷3,CSR
[ [ rptblocal sample_loop-1 ; Start the outer loop
mov *AR0÷,*AR3 ; Put the new sample to signal buffer
mpym *AR3÷,*AR1+,AC0 ; Do the 1st operation
[ [ rpt CSR ; Start the inner loop
macm *AR3÷,*AR1÷,AC0
macmr *AR3,*AR1÷,AC0 ; Do the last operation
mov hi(AC0),*AR2÷ ; Save result in Q15 format
sample_loop
Four auxiliary registers, AR0±AR3, are used as pointers in this example. AR0 points
to the input buffer in[]. The signal buffer x[]containing the current input x(n) and the
L ÷1 old samples is pointed at by AR3. The filter coefficients in the array h[]are
pointed at by AR1. For each iteration, a new sample is placed into the signal buffer and
EXPERIMENTS USING THE TMS320C55X 227
the inner loop repeats the multiply±accumulate instructions. Finally, the filter output
y(n) is rounded and stored in the output buffer out[]that is pointed at by AR2.
Both AR1 and AR3 use circular addressing mode. At the end of the computation,
the coefficient pointer AR1 will be wrapped around, thus pointing at the first co-
efficient again. The signal buffer pointer AR3 will point at the oldest sample,
x(n ÷L ÷1). In the next iteration, AR1 will start from the first tap, while the oldest
sample in the signal buffer will be replaced with the new input sample as shown in
Figure 5.12.
In the assembly code, we initialize the repeat counter CSR with the value L-3 for L-2
iterations. This is because we use a multiplication instruction before the loop and a
multiply±accumulate-rounding instruction after the repeat loop. Moving instructions
outside the repeat loops is called loop unrolling. It is clear from using the loop unrolling
technique, the FIR filter must have at least three coefficients. The complete assembly
program fir.asm is given in the experimental software package.
The C program exp5a.c listed in Table 5.2 will be used for Experiment 5A. It uses
the data file input5.dat as input, and calls fir()to perform lowpass filtering. Since
we use circular buffers, we define a global variable index as the signal buffer index for
tracking the starting position of the signal buffer for each sample block. The C55x
compiler supports several pragma directives. We apply the two most frequently used
directives, CODE_SECTION and DATA_SECTION, to allocate the C functions' program
and data variables for experiments. For a complete list of C pragma directives that the
compiler supports, please refer to the TMS320C55x Optimizing C Compiler User's
Guide [11].
Table 5.2 List of the C program for Experiment 5A
/*
exp5a.c ÷ Block FIR filter experiment using input data file
*/
#define M 128 /* Input sample size */
#define L 48 /* FIR filter order */
#define SN L /* Signal buffer size */
extern unsigned int fir(int *, unsigned int, int *, unsigned int,
int *, int *, unsigned int);
/* Define DSP system memory map */
#pragma DATA_SECTION(LP_h, "fir_coef");
#pragma DATA_SECTION(x, "fir_data");
#pragma DATA_SECTION(index, "fir_data");
#pragma DATA_SECTION(out, "output");
#pragma DATA_SECTION(in, "input");
#pragma DATA_SECTION(input, "input");
#pragma CODE_SECTION(main, "fir_code");
/* Input data */
#include "input5.dat"
228 DESIGN AND IMPLEMENTATION OF FIR FILTERS
Table 5.2 (continued )
/* Low-pass FIR filter coefficients */
int LP_h[L]=
{÷6,28,46,27,÷35,÷100,÷93,26,191,240,52,÷291,÷497,÷278,
337,888,773,÷210,÷1486,÷1895,÷442,2870,6793,9445,
9445,6793,2870,÷442,÷1895,÷1486,÷210,773,888,337,
÷278,÷497,÷291,52,240,191,26,÷93,÷100,÷35,27,46,28,÷6};
int x[SN]; /* Signal buffer */
unsigned int index; /* Signal buffer index */
int out[M]; /* Output buffer */
int in[M]; /* Input buffer */
void main(void)
{
unsigned int i,j;
/* Initialize filter signal buffer */
for(i = 0; i < SN;i÷÷)
x[i]= 0;
index = 0;
/* Processing samples using a block FIR filter */
j = 0;
for(;;)
{
for(i = 0; i < M; i÷÷)
{
in[i]= input [j÷÷]; /* Get a buffer of samples */
if(j == 160)
j = 0;
}
index = fir(in,M,LP_h,L,out,x,index); /* FIR filtering */
}
}
The CODE_SECTION is a pragma directive that allocates code objects to a named
code section. The syntax of the pragma is given as:
#pragma CODE_SECTION(func_name, "section_name");
where the func_name is the name of the C function that will be allocated into the
program memory section defined by the section_name. The linker command file uses
this section to define and allocate a specific memory section for the C function.
The DATA_SECTION is the pragma directive that allocates data objects to a named
data section. The syntax of the pragma is given as:
#pragma DATA_SECTION(var_name, "section_name");
EXPERIMENTS USING THE TMS320C55X 229
where the var_name is the variable name contained in the C function that will be
allocated into the data section defined by the section_name. The linker command file
uses this name for data section allocation to system memory.
Go through the following steps for Experiment 5A:
1. Copy the assembly program fir.asm, the C function epx5a.c, the linker com-
mand file exp5.cmd, and the experimental data input5.dat from the software
package to the working directory.
2. Create the project exp5a and add the files fir.asm, epx5a.c, and exp5.cmd to
the project.
2. The prototype of the FIR routine is defined as:
unsigned int fir(int *in, unsigned int M, int *h,
unsigned int L, int *out, int *x, int index);
2. where in is the pointer to the input data buffer in[], M defines the number of
samples in the input data buffer in[], h is the pointer to the filter coefficient buffer
LP_h[], L is the number of FIR filter coefficients, out is the pointer to the output
buffer out[], x is the pointer to the signal buffer x[], and index is the index for
signal buffer.
3. Build, debug, and run the project exp5a. The lowpass filter will attenuate two
higher frequency components at 1800 and 3300 Hz and pass the 800 Hz sinewave.
Figure 5.25 shows the time-domain and frequency-domain plots of the input and
output signals.
4. Use the CCS animation capability to view the FIR filtering process frame by frame.
Profile the FIR filter to record the memory usage and the average C55x cycles used
for processing one block of 128 samples.
5.6.2 Experiment 5B ± Implementation of Symmetric FIR Filter
As shown in Figure 5.7, a symmetric FIR filter has the characteristics of symmetric
impulse responses (or coefficients) about its center index. Type I FIR filters have an
even number of symmetric coefficients, while Type II filters have an odd number. An
even symmetric FIR filter shown in Figure 5.8 indicates that only the first half of the
filter coefficients are necessary for computing the filter result.
The TMS320C55x has two special instructions, firsadd and firssub, for imple-
menting symmetric and anti-symmetric FIR filters. The former can be used to compute
symmetric FIR filters given in (5.2.23), while the latter can be used for anti-symmetric
FIR filters defined in (5.2.24). The syntax for symmetric and anti-symmetric filter
instructions are
firsadd Xmem,Ymem,Cmem,ACx,ACy ; Symmetric FIR filter
firssub Xmem,Ymem,Cmem,ACx,ACy ; Anti-symmetric FIR filter
230 DESIGN AND IMPLEMENTATION OF FIR FILTERS
(a) (b)
Figure 5.25 Input and output signals of Experiment 5A: (a) input signals in the frequency (top)
and time (bottom) domains, and (b) output signals in the frequency (top) and time (bottom)
domains
where Xmem and Ymem are the signal buffers for ¦x(n), x(n ÷1), . . . x(n ÷L=2 ÷1)¦
and ¦x(n ÷L=2), . . . x(n ÷L ÷1)¦, and Cmem is the coefficient buffer.
For a symmetric FIR filter, the firsadd instruction is equivalent to performing the
following parallel instructions in one cycle:
macm *CDP÷,ACx,ACy ; b
l
[x(n÷l) ÷ x(n ÷ l ÷ L ÷1)]
[ [ add *ARx÷,*ARy÷,ACx ; x (n ÷ l ÷1) ÷ x(n ÷ l ÷ L ÷ 2)
While the macm instruction carries out the multiply±accumulate portion of the sym-
metric filter operation, the add instruction adds up a pair of samples for the next
iteration. This parallel arrangement effectively improves the computation of symmetric
FIR filters. The following assembly program shows an implementation of symmetric
FIR filter using the TMS320C55x:
mov #M÷1,BRC0 ; Outer loop counter for execution
mov #(L/2÷3),CSR ; Inner loop for (L/2÷2) iteration
mov #L/2,T0 ; Set up pointer offset for AR1
sub #(L/2÷2),T1 ; Set up pointer offset for AR3
[ [ rptblocal sample_loop-1 ; To prevent overflow in addition
mov *AR0÷,AC1 ; Get new sample
mov #0,AC0 ; Input is scaled to Q14 format
[ [ mov AC1<#÷1,*AR3 ; Put input to signal buffer
add *AR3÷,*AR1÷,AC1 ; AC1 = [x(n)÷x(n÷L÷1)]< 16
[ [ rpt CSR ; Do L/2÷2 iterations
EXPERIMENTS USING THE TMS320C55X 231
firsadd *AR3÷,*AR1÷,*CDP÷,AC1,AC0
firsadd *(AR3÷T0),*(AR1÷T1),*CDP÷,AC1,AC0
macm *CDP÷,AC1,AC0 ; Finish the last macm instruction
mov rnd(hi(AC0<1)),*AR2÷; Save rounded & scaled result
sample_loop
Although the assembly program of the symmetric FIR filter is similar to the regular
FIR filter, there are several differences when we implement it using the firsadd
instruction: (1) We only need to store the first half of the symmetric FIR filter coeffi-
cients. (2) The inner-repeat loop is set to L=2÷2 since each multiply±accumulate
operation accounts for a pair of samples. (3) In order to use firsadd instructions
inside a repeat loop, we add the first pair of filter samples using a dual-memory add
instruction, add *AR1÷,*AR3÷,AC1. We also place the last instruction, macmr
*CDP÷, AC1, AC0, outside the repeat loop for the final calculation. (4) We use
two data pointers, AR1 and AR3, to address the signal buffer. AR3 points at the
newest sample in the buffer, and AR1 points at the oldest sample in the buffer.
Temporary registers, T1 and T0, are used as the offsets for updating circular buffer
pointers. The offsets are initialized to T0 = L/2 and T1 = L=2÷2. After AR3 and AR1
are updated, they will point to the newest and the oldest samples again. Figure 5.26
illustrates this two-pointer circular buffer for a symmetric FIR filtering. The firsadd
instruction accesses three data buses simultaneously (Xmem and Ymem for signal samples
and Cmem for filter coefficient). The coefficient pointer CDP is set as the circular
pointer for coefficients. The input and output samples are pointed at by AR0 and AR2.
Two implementation issues should be considered: (1) The symmetric FIR filtering
instruction firsadd adds two corresponding samples, and then performs multiplica-
tion. The addition may cause an undesired overflow. (2) The firsadd instruction
accesses three read operations in the same cycle. This may cause data memory bus
contention. The first problem can be resolved by scaling the new sample to Q14 format
x(n − L + 2)
x(n) x(n − L + 1)
x(n − 1)
x(n − 2)
x(n − 3)
AR3 at time n AR1 at time n
(a)
x(n − L + 2)
x(n) x(n − L + 1)
x(n − 1)
x(n − 2)
x(n − 3)
AR1 for next x(n − L + 1)
AR3 for next x(n)
(b)
Figure 5.26 Circular buffer for accessing signals for a symmetric FIR filtering. The pointers to
x(n) and x(n ÷L ÷1) are updated at the counter-clockwise direction: (a) circular buffer for a
symmetric FIR filter at time n, and (b) circular buffer for a symmetric FIR filter at time n ÷1
232 DESIGN AND IMPLEMENTATION OF FIR FILTERS
prior to saving it to the signal buffer. The filter result needs to be scaled back before it
can be stored into the output buffer. The second problem can be resolved by placing the
filter coefficient buffer and the signal buffer into different memory blocks.
Go through the following steps for Experiment 5B:
1. Copy the assembly programfirsymm.asm, the Cfunction epx5b.c, and the linker
command file exp5.cmd from the software package into the working directory.
2. Create the project exp5b and add the files firsymm.asm, epx5b.c, and
exp5.cmd into the project.
3. The symmetric FIR routine is defined as:
unsigned int firsymm(int *in, unsigned int M, int *h, unsigned int L,
int *out, int *x, unsigned int index);
4. where all the arguments are the same as those defined by fir() in Experiment 5A.
4. Build and run the project exp5b. Compare the results with Figure 5.25.
5. Profile the symmetric filter performance and record the memory usage. How many
instruction cycles were reduced as a result of using the symmetric FIR filter
implementation? How many memory locations have been saved?
5.6.3 Experiment 5C ± Implementation of FIR Filter Using
Dual-MAC
As introduced in Chapter 2, the TMS320C55x has two multiply±accumulate (MAC)
units and four accumulators. We can take advantage of the dual-MAC architecture to
improve the execution speed of FIR filters. Unlike the symmetric FIR filter implemen-
tations given in the previous experiment, FIR filters implemented using dual-MAC can
be symmetric, anti-symmetric, or any other type. The basic idea of using dual-MAC
to improve the processing speed is to generate two outputs in parallel. That is, we use
one MAC unit to compute y(n) and the other to generate y(n ÷1) at the same time.
For example, the following parallel instructions can be used for dual-MAC filtering
operations:
rpt CSR
mac *ARx+,*CDP÷,ACx ; ACx ÷= bl*x(n)
:: mac *ARy÷,*CDP÷,ACy ; ACy ÷= bl*x(n÷1)
In this example, ARx and ARy are data pointers to x(n) and x(n ÷1), respectively, and
CDP is the coefficient pointer. The repeat loop produces two filter outputs, y(n) and
y(n ÷1). After execution, the addresses of the data pointers ARx and ARy are increased
by one. The coefficient pointer CDP is also incremented by one, although the coefficient
pointer CDP is set for auto-increment mode in both instructions. This is because when
EXPERIMENTS USING THE TMS320C55X 233
CDP pointer is used in parallel instructions, it can be incremented only once. Figure
5.27 shows the C55x dual-MAC architecture for FIR filtering. The CDP uses B-bus to
fetch filter coefficients, while ARx and ARy use C-bus and D-bus to get data from the
signal buffer. The dual-MAC filtering results are temporarily stored in the accumulators
ACx and ACy.
The following example shows the C55x assembly implementation using the dual-
MAC and circular buffer for a block-FIR filter:
mov #M÷1,BRC0 ; Outer loop counter
mov #(L/2÷3),CSR ; Inner loop counter as L/2÷2
[ [ rptblocal sample_loop-1
mov *AR0÷,*AR1 ; Put new sample to signal buffer x[n]
mov *AR0÷,*AR3 ; Put next new sample to location x[n÷1]
mpy *AR1÷,*CDP÷,AC0 ; First operation
:: mpy *AR3÷,*CDP÷,AC1
[ [ rpt CSR
mac *AR1÷,*CDP÷,AC0 ; Rest MAC iterations
:: mac *AR3÷,*CDP÷,AC1
macr *AR1,*CDP÷,AC0
:: macr *AR3,*CDP÷,AC1 ; Last MAC operation
mov pair(hi(AC0)),dbl(*AR2÷) ; Store two output data
sample_loop
There are three implementation issues to be considered: (1) In order to use dual-MAC
units, we need to increase the length of the signal buffer by one in order to accommodate
an extra memory location required for computing two output signals. With an add-
itional space in the buffer, we can form two sample sequences in the signal buffer, one
pointed at by AR1 and the other by AR3. (2) The dual-MAC implementation of the
FIRfilter also makes three memory reads simultaneously. Two memory reads are used to
get data samples from the signal buffer into MAC units, and the third one is used to fetch
the filter coefficient. To avoid memory bus contention, we shall place the coefficients
in a different memory block. (3) We place the convolution sums in two accumulators
when we use dual-MAC units. To store both filter results, it requires two memory
store instructions. It will be more efficient if we can use the dual-memory-store instruc-
tion, mov pair(hi(AC0)),dbl(*AR2÷), to save both outputs y(n) and y(n ÷1) to the
data memory in the same cycle. However, this requires the data memory to be
B-bus
C-bus
D-bus
ARx ARy
MAC MAC
ACy ACx
CDP
Figure 5.27 Diagram of TMS320C55x dual-MAC architecture
234 DESIGN AND IMPLEMENTATION OF FIR FILTERS
aligned on an even word boundary. This alignment can be done using the linker
command file with the key word, align 4, see the linker command file. We use the
DATA_SECTION pragma directive to tell the linker where to place the output
sequence.
Go through the following steps for Experiment 5C:
1. Copy the assembly program firs2macs.asm, the C function epx5c.c, and
the linker command file exp5.cmd from the software package to the working
directory. The dual-MAC FIR routine is defined as:
unsigned int fir2macs(int *in, unsigned int M, int *h, unsigned int L,
int *out, int *x, unsigned int index);
1. where all the arguments are the same as Experiment 5A.
2. Create the project exp5c and add files fir2macs.asm, epx5c.c, and exp5.cmd
into the project.
3. Build and run the project. Compare the results with the results from the two
previous experiments.
4. Profile the filter performance and record the memory usage. How many instruction
cycles were reduced by using the dual-MAC implementation? Why is the dual-MAC
implementation more efficient than the symmetric FIR implementation?
References
[1] N. Ahmed and T. Natarajan, Discrete-Time Signals and Systems, Englewood Cliffs, NJ: Prentice-
Hall, 1983.
[2] V. K. Ingle and J. G. Proakis, Digital Signal Processing Using MATLAB V.4, Boston: PWS
Publishing, 1997.
[3] Signal Processing Toolbox for Use with MATLAB, Math Works, 1994.
[4] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, Englewood Cliffs, NJ:
Prentice-Hall, 1989.
[5] S. J. Orfanidis, Introduction to Signal Processing, Englewood Cliffs, NJ: Prentice-Hall, 1996.
[6] J. G. Proakis and D. G. Manolakis, Digital Signal Processing ± Principles, Algorithms, and
Applications, 3rd Ed., Englewood Cliffs, NJ: Prentice-Hall, 1996.
[7] S. K. Mitra, Digital Signal Processing: A Computer-Based Approach, 2nd Ed., New York, NY:
McGraw Hill, 1998.
[8] D. Grover and J. R. Deller, Digital Signal Processing and the Microcontroller, Englewood Cliffs,
NJ: Prentice-Hall, 1999.
[9] F. Taylor and J. Mellott, Hands-On Digital Signal Processing, New York, NY: McGraw Hill,
1998.
[10] S. D. Stearns and D. R. Hush, Digital Signal Analysis, 2nd Ed., Englewood Cliffs, NJ: Prentice-
Hall, 1990.
[11] Texas Instruments, Inc., TMS320C55x Optimizing C Compiler User's Guide, Literature no.
SPRU281, 2000.
REFERENCES 235
Exercises
Part A
1. Consider the moving-average filter given in Example 5.4. What is the 3-dB bandwidth of this
filter if the sampling rate is 8 kHz?
2. Consider the FIR filter with the impulse response h(n) = ¦1, 1, 1¦. Calculate the magnitude
and phase responses and show that the filter has linear phase.
3. Given a linear time-invariant filter with an exponential impulse response
h(n) = a
n
u(n),
show that the output due to a unit-step input, u(n), is
y(n) =
1 ÷a
n÷1
1 ÷a
, n _ 0:
4. A rectangular function of length L can be expressed as
x(n) = u(n) ÷u(n ÷L) =
1, 0 _ n _ L ÷1
0, elsewhere,

show that
(a) r(n) = u(n) + u(n), where + denotes linear convolution and r(n) = (n ÷1)u(n) is called the
unit-ramp sequence.
(b) t(n) = x(n) + x(n) = r(n) ÷2r(n ÷L) ÷r(n ÷2L) is the triangular pulse.
5. Using the graphical interpretation of linear convolution given in Figure 5.4 to compute the
linear convolution of h(n) = ¦1, 2, 1¦ and x(n), n = 0, 1, 2 defined as follows:
(a) x(n) = ¦1, ÷1, 2¦,
(b) x(n) = ¦1, 2, ÷1¦, and
(c) x(n) = ¦1, 3, 1¦:
6. The comb filter also can be described as
y(n) = x(n) ÷x(n ÷L),
find the transfer function, zeros, and the magnitude response of this filter and compare the
results with Figure 5.6.
7. Show that at ! = 0, the magnitude of the rectangular window function is 2M÷1.
8. Assuming h(n) has the symmetry property h(n) = h(÷n) for n = 0, 1, . . . M, show that H(!)
can be expressed as
236 DESIGN AND IMPLEMENTATION OF FIR FILTERS
H(!) = h(0) ÷
¸
M
n=1
2h(n) cos(!n):
9. The simplest digital approximation to a continuous-time differentiator is the first-order
operation defined as
y(n) =
1
T
[x(n) ÷x(n ÷1)[:
Find the transfer function H(z), the frequency response H(!), and the phase response of the
differentiator.
10. Redraw the signal-flow diagram shown in Figure 5.8 and modify equations (5.2.23) and
(5.2.24) in the case that L is an odd number.
11. Consider the rectangular window w(n) of length L = 2M÷1 defined in (5.3.17). Show that
the convolution of w(n) with itself and then divided by L yields the triangular window.
12. Assuming that H(!) given in (5.3.2) is an even function in the interval [![ < p, show that
h(n) =
1
p

p
0
H(!) cos(!n)d!, n _ 0
and h(÷n) = h(n).
13. Design a lowpass FIR filter of length L = 5 with a linear phase to approximate the ideal
lowpass filter of cut-off frequency !
c
= 1. Use the Hamming window to eliminate the ripples
in the magnitude response.
14. The ideal highpass filter of Figure 5.1(b) has frequency response
H(!) =
0, [![ < !
c
1, !
c
_ [![ _ p.

Compute the coefficients of a causal FIR filter of length L = 2M÷1 obtained by truncating
and shifting the impulse response of the ideal bandpass filter.
15. Design a bandpass filter
H( f ) =
1, 1:6 kHz _ f _ 2 kHz
0, otherwise

with the sampling rate 8 kHz and the duration of impulse response be 50 msec using Fourier
series method.
Part B
16. Consider the FIR filters with the following impulse responses:
(a) h(n) = ¦÷4, 1, ÷1, ÷2, 5, 0, ÷5, 2, 1, ÷1, 4¦
EXERCISES 237
(b) h(n) = ¦÷4, 1, ÷1, ÷2, 5, 6, 5, ÷2, ÷1, 1, ÷4¦
Using MATLAB to plot magnitude responses, phase responses, and locations of zeros of the
FIR filter's transfer function H(z).
17. Show the frequency response of the lowpass filter given in (5.2.10) for L = 8 and compare
the result with Figure 5.6.
18. Plot the magnitude response of a linear-phase FIR highpass filter of cut-off frequency
!
c
= 0:6p by truncating the impulse response of the ideal highpass filter to length
L = 2M÷1 for M = 32 and 64.
19. Repeat problem 18 using Hamming and Blackman window functions. Show that oscillatory
behavior is reduced using the windowed Fourier series method.
20. Write C (or MATLAB) program that implement a comb filter of L = 8. The program must
have the input/output capability as introduced in Appendix C. Test the filter using the
sinusoidal signals of frequencies and !
1
= p=4 and !
2
= 3p=8. Explain the results based
on the distribution of the zeros of the filter.
21. Rewrite the above program using the circular buffer.
22. Rewrite the program firfltr.c given in Appendix C using circular buffer. Implement the
circular pointer updated in a new C function to replace the function shift.c.
Part C
23. Based on the assembly routines given in Experiments 5A, 5B, and 5C, what is the minimum
number of the FIR filter coefficients if the FIR filter is
(a) symmetric and L is even,
(b) symmetric and L is odd,
(c) anti-symmetric and L is even,
(d) anti-symmetric and L is odd.
Do we need to modify these routines if the FIR filter has odd number of taps?
24. Design a 24th-order bandpass FIR filter using MATLAB. The filter will attenuate the
800 Hz and 3.3 kHz frequency components of the signal generated by the signal generator
signal_gen(). Implement this filter using the C55x assembly routines fir.asm,
firsymm.asm, and fir2macs.asm. Plot the filter results in both the time domain and
the frequency domain.
25. When design highpass or bandstop FIR filters using MATLAB, the number of filter
coefficients is an odd number. This ensures the unit gain at the half-sampling frequency.
Design a highpass FIR filter, such that it will pass the 3.3 kHz frequency components of the
input signal. Implement this filter using the dual-MAC block FIR filter. Plot the results in
238 DESIGN AND IMPLEMENTATION OF FIR FILTERS
both the time domain and the frequency domain (Hint: modify the assembly routine
fir2macs.asm to handle the odd number coefficients).
26. Design an anti-symmetric bandpass FIR filter to allow only the frequency component at
1.8 kHz to pass. Using firssub instruction to implement the FIR filter and plot the filter
results in both the time domain and the frequency domain.
27. Experiment 5B demonstrates a symmetric FIR filter implementation. This filter can also be
implemented efficiently using the C55x dual-MAC architecture. Modify the dual-MAC FIR
filter assembly routine fir2macs.asm to implement the Experiment 5B based on the
Equation (5.2.23). Compare the profiling results with Experiment 5B that uses the symmetric
FIR filter firsadd instruction.
28. Use TMS320C55x EVM (or DSK) for collecting real-time signal from an analog signal
generator.
± set the TMS320C55x EVM or DSK to 8 kHz sampling rate
± connect the signal generator output to the audio input of the EVM/DSK
± write an interrupt service routine (ISR) to handle input samples
± process the samples at 128 samples per block
± verify your result using an oscilloscope or spectrum analyzer
EXERCISES 239
6
Design and Implementation
of IIR Filters
We have discussed the design and implementation of digital FIR filters in the previous
chapter. In this chapter, our attention will be focused on the design, realization, and
implementation of digital IIR filters. The design of IIR filters is to determine the
transfer function H(z) that satisfies the given specifications. We will discuss the basic
characteristics of digital IIR filters, and familiarize ourselves with the fundamental
techniques used for the design and implementation of these filters. IIR filters have the
best roll-off and lower sidelobes in the stopband for the smallest number of coefficients.
Digital IIR filters can be easily obtained by beginning with the design of an analog
filter, and then using mapping technique to transform it from the s-plane into the z-
plane. The Laplace transform will be introduced in Section 6.1 and the analog filter will
be discussed in Section 6.2. The impulse-invariant and bilinear-transform methods for
designing digital IIR filters will be introduced in Section 6.3, and realization of IIR
filters using direct, cascade, and parallel forms will be introduced in Section 6.4. The
filter design using MATLAB will be described in Section 6.5, and the implementation
considerations are given in Section 6.6. The software development and experiments
using the TMS320C55x will be given in Section 6.7.
6.1 Laplace Transform
As discussed in Chapter 4, the Laplace transform is the most powerful technique used to
describe, represent, and analyze analog signals and systems. In order to introduce
analog filters in the next section, a brief review of the Laplace transform is given in
this section.
6.1.1 Introduction to the Laplace Transform
Many practical aperiodic functions such as a unit step function u(t), a unit ramp tu(t),
or an impulse train
¸
·
k=÷·
d(t ÷kT) do not satisfy the integrable condition given in
(4.1.11), which is a sufficient condition for a function x(t) that possesses a Fourier
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
transform. Given a positive-time function, x(t) = 0, for t < 0, a simple way to find the
Fourier transform is to multiply x(t) by a convergence factor e
÷st
, where s is a positive
number such that

·
0

x(t)e
÷st

dt < ·: (6:1:1)
Taking the Fourier transform defined in (4.1.10) on the composite function x(t)e
÷st
, we
have
X(s) =

·
0
x(t)e
÷st
e
÷jVt
dt =

·
0
x(t)e
÷(s÷jV)t
dt
=

·
0
x(t)e
÷st
dt, (6:1:2)
where
s = s ÷jV (6:1:3)
is a complex variable. This is called the one-sided Laplace transform of x(t) and is
denoted by X(s) = LT[x(t)[. Table 6.1 lists the Laplace transforms of some simple time
functions.
Example 6.1: Find the Laplace transform of signal
x(t) = a ÷be
÷ct
, t _ 0:
From Table 6.1, we have the transform pairs
a ÷
a
s
and e
÷ct
÷
1
s ÷c
:
Using the linear property, we have
X(s) =
a
s
÷
b
s ÷c
:
The inverse Laplace transform can be expressed as
x(t) =
1
2pj

s÷j·
s÷j·
X(s)e
st
ds: (6:1:4)
The integral is evaluated along the straight line s ÷jV in the complex plane from
V = ÷· to V = ·, which is parallel to the imaginary axis jV at a distance s
from it.
242 DESIGN AND IMPLEMENTATION OF IIR FILTERS
Table 6.1 Basic Laplace transform pairs
x(t), t _ 0 X(s)
d(t) 1
u(t)
1
s
c
c
s
ct
c
s
2
ct
n÷1
c(n ÷1)!
s
n
e
÷at
1
s ÷a
sinV
0
t
V
0
s
2
÷V
2
0
cos V
0
t
s
s
2
÷V
2
0
x(t) cos V
0
t
1
2
[X(s ÷jV
0
) ÷X(s ÷jV
0
)[
x(t) sin V
0
t
j
2
[X(s ÷jV
0
) ÷X(s ÷jV
0
)[
e
±at
x(t) X(s
÷
÷
a)
x(at)
1
a
X
s
a

Equation (6.1.2) clearly shows that the Laplace transform is actually the Fourier
transform of the function x(t)e
÷st
, t > 0. From (6.1.3), we can think of a complex s-
plane with a real axis s and an imaginary axis jV. For values of s along the jV axis, i.e.,
s = 0, we have
X(s)[
s=jV
=

·
0
x(t)e
÷jVt
dt, (6:1:5)
which is the Fourier transform of the causal signal x(t). Given a function X(s), we can
find its frequency characteristics by setting s = jV.
There are convolution properties associated with the Laplace transform. If
y(t) = x(t) + h(t) =

·
0
x(t)h(t ÷t)dt =

·
0
h(t)x(t ÷t)dt, (6:1:6)
LAPLACE TRANSFORM 243
then
Y(s) = X(s)H(s), (6:1:7)
where Y(s), H(s), and X(s) are the Laplace transforms of y(t), h(t), and x(t), respectively.
Thus convolution in the time domain is equivalent to multiplication in the Laplace (or
frequency) domain.
In (6.1.7), H(s) is the transfer function of the system defined as
H(s) =
Y(s)
X(s)
=

·
0
h(t)e
÷st
dt, (6:1:8)
where h(t) is the impulse response of the system. The general form of a transfer function
is expressed as
H(s) =
b
0
÷b
1
s ÷ ÷b
L÷1
s
L÷1
a
0
÷a
1
s ÷ ÷a
M
s
M
=
N(s)
D(s)
: (6:1:9)
The roots of N(s) are the zeros of the transfer function H(s), while the roots of D(s) are
the poles.
Example 6.2: The input signal x(t) = e
÷2t
u(t) is applied to an LTI system, and the
output of the system is given as
y(t) = (e
÷t
÷e
÷2t
÷e
÷3t
)u(t):
Find the system's transfer function H(s) and the impulse response h(t).
From Table 6.1, we have
X(s) =
1
s ÷2
and Y(s) =
1
s ÷1
÷
1
s ÷2
÷
1
s ÷3
:
From (6.1.8), we obtain
H(s) =
Y(s)
X(s)
= 1 ÷
s ÷2
s ÷1
÷
s ÷2
s ÷3
:
This transfer function can be written as
H(s) =
s
2
÷6s ÷7
(s ÷1)(s ÷3)
= 1 ÷
1
s ÷1
÷
1
s ÷3
:
From Table 6.1, we have
h(t) = d(t) ÷(e
÷t
÷e
÷3t
)u(t):
244 DESIGN AND IMPLEMENTATION OF IIR FILTERS
The stability condition for a system can be represented in terms of its impulse
response h(t) or its transfer function H(s). A system is stable if
lim
t÷·
h(t) = 0: (6:1:10)
This condition is equivalent to requiring that all the poles of H(s) must be in the left-half
of the s-plane, i.e., s < 0.
Example 6.3: Consider the impulse response
h(t) = e
÷at
u(t):
This function satisfies (6.1.10) for a > 0. From Table 6.1, the transfer function
H(s) =
1
s ÷a
, a > 0
has the pole at s = ÷a, which is located at the left-half s-plane. Thus the system is
stable.
If lim
t÷·
h(t) ÷·, the system is unstable. This condition is equivalent to the system
that has one or more poles in the right-half s-plane, or has multiple-order pole(s) on the
jV axis. The system is marginally stable if h(t) approaches a non-zero value or a
bounded oscillation as t approaches infinity. If the system is stable, then the natural
response goes to zero as t ÷·. In this case, the natural response is also called the
transient response. If the input signal is periodic, then the corresponding forced
response is called the steady-state response. When the input signal is the sinusoidal
signal in the form of sin Vt, cos Vt, or e
jVt
, the steady-state output is called the
sinusoidal steady-state response.
6.1.2 Relationships between the Laplace and z-Transforms
An analog signal x(t) can be converted into a train of narrow pulses x(nT) as
x(nT) = x(t)d
T
(t), (6:1:11)
where
d
T
(t) =
¸
·
n=÷·
d(t ÷nT) (6:1:12)
represents a unit impulse train and is called a sampling function. Clearly, d
T
(t) is not
a signal that we could generate physically, but it is a useful mathematical abstrac-
tion when dealing with discrete-time signals. Assuming that x(t) = 0 for t < 0, we
have
LAPLACE TRANSFORM 245
x(nT) = x(t)
¸
·
n=÷·
d(t ÷nT) =
¸
·
n=0
x(nT)d(t ÷nT): (6:1:13)
To obtain the frequency characteristics of the sampled signal, take the Laplace
transform of x(nT) given in (6.1.13). Integrating term-by-term and using the property
of the impulse function
¸
·
÷·
x(t)d(t ÷t)dt = x(t), we obtain
X(s) =

·
÷·
¸
·
n=0
x(nT)d(t ÷nT)
¸ ¸
e
st
dt =
¸
·
n=0
x(nT)e
÷nsT
: (6:1:14)
When defining a complex variable
z = e
sT
, (6:1:15)
Equation (6.1.14) can be expressed as
X(z) = X(s)[
z=e
sT =
¸
·
n=0
x(nT)z
÷n
, (6:1:16)
where X(z) is the z-transform of the discrete-time signal x(nT). Thus the z-transform can
be viewed as the Laplace transform of the sampled function x(t) with the change of
variable z = e
sT
.
As discussed in Chapter 4, the Fourier transform of a sequence x(nT) can be obtained
from the z-transform by replacing z with e
j!
. That is, by evaluating the z-transform on
the unit circle of [z[ = 1. The whole procedure can be summarized in Figure 6.1.
6.1.3 Mapping Properties
The relationship z = e
sT
defined in (6.1.15) represents the mapping of a region in the s-
plane to the z-plane since both s and z are complex variables. Since s = s ÷jV, we have
z = e
sT
= e
sT
e
jVT
= [z[e
j!
, (6:1:17)
Sampling
Laplace
transform
z-transform
Fourier
transform
z = e
sT
z = e
jw
x(t) x(nT)
X(s) X(z) X(w)
Figure 6.1 Relationships between the Laplace, Fourier, and z-transforms
246 DESIGN AND IMPLEMENTATION OF IIR FILTERS
w = p w = 0
w = p/ 2
jΩ
s = 0
−p/T
p/T
s
s > 0
Im z
Re z
|z| = 1
w = 3p/ 2
s-plane z-plane
s < 0
Figure 6.2 Mapping between the s-plane and z-plane
where the magnitude
[z[ = e
sT
(6:1:18a)
and the angle
! = VT: (6:1:18b)
When s = 0, the amplitude given in (6.1.18a) is [z[ = 1, and Equation (6.1.17) is
simplified to z = e
jVT
. It is apparent that the portion of the jV-axis between
V = ÷p=T and V = p=T in the s-plane is mapped onto the unit circle in the z-plane
from ÷p to p as illustrated in Figure 6.2. As V increases from p=T to 3p=T in the
s-plane, another counterclockwise encirclement of the unit circle results in the z-plane.
Thus as V varies from 0 to ·, there are an infinite number of encirclements of the unit
circle in the counterclockwise direction. Similarly, there are an infinite numbers of
encirclements of the unit circle in the clockwise direction as V varies from 0 to ÷·.
From (6.1.18a), [z[ < 1 when s < 0. Thus each strip of width 2p=T in the left-half of
the s-plane is mapped onto the unit circle. This mapping occurs in the form of con-
centric circles in the z-plane as s varies from 0 to ÷·. Equation (6.1.18a) also implies
that [z[ > 1 if s > 0. Thus each strip of width 2p=T in the right-half of the s-plane is
mapped outside of the unit circle. This mapping also occurs in concentric circles in the z-
plane as s varies from 0 to ·.
In conclusion, the mapping from the s-plane to the z-plane is not one-to-one, since
there is more than one point in the s-plane that corresponds to a single point in the z-
plane. This issue will be discussed later when we design a digital filter from a given
analog filter.
6.2 Analog Filters
Analog filter design is a well-developed technique. Many of the techniques employed in
studying digital filters are analogous to those used in studying analog filters. The most
systematic approach to designing IIR filters is based on obtaining a suitable analog
filter function and then transforming it into the discrete-time domain. This is not
possible when designing FIR filters as they have no analog counterpart.
ANALOG FILTERS 247
6.2.1 Introduction to Analog Filters
In this section, we briefly introduce some basic concepts of analog filters. Knowledge of
analog filter transfer functions is readily available since analog filters have already been
investigated in great detail. In Section 6.3, we will introduce a conventional powerful
bilinear-transform method to design digital IIR filters utilizing analog filters.
From basic circuit theory, capacitors and inductors have an impedance (X) that
depends on frequency. It can be expressed as
X
C
=
1
jVC
(6:2:1)
and
X
L
= jVL, (6:2:2)
where C is the capacitance with units in Farads (F), and L is the inductance with units
in Henrys (H). When either component is combined with a resistor, we can build
frequency-dependent voltage dividers. In general, capacitors and resistors are used to
design analog filters since inductors are bulky, more expensive, and do not perform as
well as capacitors.
Example 6.4: Consider a circuit containing a resistor and a capacitor as shown in
Figure 6.3. Applying Ohm's law to this circuit, we have
V
in
= I(R ÷X
C
) and V
out
= IR:
From (6.2.1), the transfer function of the circuit is
H(V) =
V
out
V
in
=
R
R ÷
1
jVC
=
jVRC
1 ÷jVRC
: (6:2:3)
The magnitude response of circuit can be expressed as
[H(V)[ =
R

R
2
÷
1
V
2
C
2
:
I
C
V
in
X(s)
V
out
Y(s)
R
Figure 6.3 An analog filter with a capacitor and resistor
248 DESIGN AND IMPLEMENTATION OF IIR FILTERS
0

H(Ω)
1
Figure 6.4 Amplitude response of analog circuit shown in Figure 6.3
The plot of the magnitude response [H(V)[ vs. the frequency V is shown in Figure
6.4. For a constant input voltage, the output is approximately equal to the input at
high frequencies, and the output approaches zero at low frequencies. Therefore
the circuit shown in Figure 6.3 is called a highpass filter since it only allows high
frequencies to pass without attenuation.
The transfer function of the circuit shown in Figure 6.3 is given by
H(s) =
Y(s)
X(s)
=
R
R ÷1=Cs
=
RCs
1 ÷RCs
: (6:2:4)
To design an analog filter, we can use computer programs to calculate the correct values
of the resistor and the capacitor for desired magnitude and phase responses. Unfortu-
nately, the characteristics of the components drift with temperature and time. It is
sometimes necessary to re-tune the circuit while it is being used.
6.2.2 Characteristics of Analog Filters
In this section, we briefly describe some important characteristics of commonly used
analog filters based on lowpass filters. We will discuss frequency transformations for
converting a lowpass filter to highpass, bandpass, and bandstop filters in the next
section. Approximations to the ideal lowpass prototype are obtained by first finding a
polynomial approximation to the desired squared magnitude [H(V)[
2
, and then con-
verting this polynomial into a rational function. An error criterion is selected to measure
how close the function obtained is to the desired function. These approximations to the
ideal prototype will be discussed briefly based on Butterworth filters, Chebyshev filters
type I and II, elliptic filters, and Bessel filters.
The lowpass Butterworth filter is an all-pole approximation to the ideal filter, which is
characterized by the squared magnitude response
[H(V)[
2
=
1
1 ÷

V

V
p

2L
, (6:2:5)
where L is the order of the filter. It is shown that [H(0)[ = 1 and [H(V
p
)[ = 1=

2

or,
equivalently, 20 log
10
[H(V
p
)[ = ÷3 dB for all values of L. Thus V
p
is called the 3-dB
ANALOG FILTERS 249
1 − dp
|H(Ω)|
1
ds
Ωp Ωs

Figure 6.5 Magnitude response of Butterworth lowpass filter
cut-off frequency. The magnitude response of a typical Butterworth lowpass filter is
illustrated in Figure 6.5. This figure shows that the magnitude is monotonically decreas-
ing in both the passband and the stopband. The Butterworth filter has a completely flat
magnitude response over the passband and the stopband. It is often referred to as the
`maximally flat' filter. This flat passband is achieved at the expense of the transition
region from V
p
to V
s
, which has a very slow roll-off. The phase response is nonlinear
around the cut-off frequency.
From the monotonic nature of the magnitude response, it is clear that the specifica-
tions are satisfied if we choose
1 _ [H(V
p
)[ _ 1 ÷d
p
, [V[ _ V
p
(6:2:6a)
in the passband, and
[H(V
s
)[ _ d
s
, [V[ _ V
s
(6:2:6b)
in the stopband. The order of the filter required to satisfy an attenuation, d
s
, at a
specified frequency, V
s
, can be determined by substituting V = V
s
into (6.2.5), resulting
in
L =
log
10
[(1 ÷d
2
s
) ÷1[
2 log
10
(V
s
=V
p
)
: (6:2:7)
The parameter L determines how closely the Butterworth characteristic approximates
the ideal filter.
If we increase the order of the filter, the flat region of the passband gets closer to the
cut-off frequency before it drops away and we have the opportunity to improve the roll-
off. Although the Butterworth filter is very easy to design, the rate at which its
magnitude decreases in the frequency range V _ V
p
is rather slow for a small L.
Therefore for a given transition band, the order of the Butterworth filter required is
often higher than that of other types of filters. In addition, for a large L, the overshoot
of the step response of a Butterworth filter is rather large.
To obtain the filter transfer function H(s), we use H(s)H(÷s)[
s=jV
= [H(V)[
2
. From
(6.2.5), the poles of the Butterworth filter are defined by
250 DESIGN AND IMPLEMENTATION OF IIR FILTERS
1 ÷(÷s
2
)
L
= 0: (6:2:8)
By solving this equation, we obtain the poles
s
k
= e
j(2k÷L÷1)p=2L
, k = 0, 1, . . . , 2L ÷1: (6:2:9)
These poles are located uniformly on a unit circle in the s-plane at intervals of p=L
radians. The pole locations are symmetrical with respect to both the real and imaginary
axes. Since 2L ÷1 cannot be an even number, it is clear that as there are no poles on the
jV axis, there are exactly L poles in each of the left- and right-half planes.
To obtain a stable Lth-order IIR filter, we choose only the poles in the left-half s-
plane. That is, we choose
s
k
= e
j(2k÷L÷1)p=2L
, k = 1, 2, . . . , L: (6:2:10)
Therefore the transfer function of Butterworth filter is defined as
H(s) =
1
(s ÷s
1
)(s ÷s
2
) . . . (s ÷s
L
)
=
1
s
L
÷a
L÷1
s
L÷1
÷. . . ÷a
1
s ÷1
: (6:2:11)
The coefficients a
k
are real numbers because the poles s
k
are symmetrical with respect to
the imaginary axis. Table 6.2 lists the denominator of the Butterworth filter transfer
function H(s) in factored form for values of L ranging from L = 1 to L = 4.
Example 6.5: Obtain the transfer function of a lowpass Butterworth filter for
L = 3. From (6.2.9), the poles are located at
s
0
= e
jp=3
, s
1
= e
j2p=3
, s
2
= e
jp
, s
3
= e
j4p=3
, s
4
= e
j5p=3
, and s
5
= 0:
These poles are shown in Figure 6.6. To obtain a stable IIR filter, we choose the
poles in the left-half plane to get
Table 6.2 Analog Butterworth lowpass filter transfer functions
L H(s)
1
1
s ÷1
2
1
s
2
÷

2

s ÷1
3
1
(s ÷1)(s
2
÷s ÷1)
4
1
(s
2
÷0:7653s ÷1)(s
2
÷1:8477s ÷1)
ANALOG FILTERS 251
H(s) =
1
(s ÷s
1
)(s ÷s
2
)(s ÷s
3
)
=
1
(s ÷e
j2p=3
)(s ÷e
jp
)(s ÷e
j4p=3
)
=
1
(s ÷1)(s
2
÷s ÷1)
:
Chebyshev filters permit a certain amount of ripples in the passband, but have a much
steeper roll-off near the cut-off frequency than what the Butterworth design can achieve.
The Chebyshev filter is called the equiripple filter because the ripples are always of equal
size throughout the passband. Even if we place very tight limits on the passband ripple,
the improvement in roll-off is considerable when compared with the Butterworth filter.
There are two types of Chebyshev filters. Type I Chebyshev filters are all-pole filters
that exhibit equiripple behavior in the passband and a monotonic characteristic in the
stopband (see Figure 6.7a). The family of type II Chebyshev filters contains both poles
and zeros, and exhibit a monotonic behavior in the passband and an equiripple behav-
ior in the stopband, as shown in Figure 6.7(b). In general, the Chebyshev filter meets the
specifications with a fewer number of poles than the corresponding Butterworth filter.
Although the Chebyshev filter is an improvement over the Butterworth filter with
respect to the roll-off, it has a poorer phase response.
The sharpest transition from passband to stopband for any given d
p
, d
s
, and L can be
achieved using the elliptic design. In fact, the elliptic filter is the optimum design in this
sense. As shown in Figure 6.8, elliptic filters exhibit equiripple behavior in both the
H(s) H(−s)
s
2
s
1 s
0
s
5
s
4
s
3
jΩ
s
Figure 6.6 Poles of the Butterworth polynomial for L = 3
(b) (a)
1 − dp
|H(Ω)|
1
ds
Ωp Ωs
Ω Ω
1 − dp
|H(Ω)|
1
ds
Ωp Ωs
Figure 6.7 Magnitude responses of Chebyshev lowpass filters: (a) type I, and (b) type II
252 DESIGN AND IMPLEMENTATION OF IIR FILTERS
1 − dp
|H(Ω)|
1
ds
Ωp Ωs

Figure 6.8 Magnitude response of elliptic lowpass filter
passband and the stopband. In addition, the phase response of elliptic filter is extremely
nonlinear in the passband (especially near cut-off frequency), so we can only use the
design where the phase is not an important design parameter.
Butterworth, Chebyshev, and elliptic filters approximate an ideal rectangular band-
width. The Butterworth filter has a monotonic magnitude response. By allowing ripples
in the passband for type I and in the stopband for type II, the Chebyshev filter can
achieve sharper cutoff with the same number of poles. An elliptic filter has even sharper
cutoffs than the Chebyshev filter for the same complexity, but it results in both pass-
band and stopband ripples. The design of these filters strives to achieve the ideal
magnitude response with trade-offs in phase response.
Bessel filters are a class of all-pole filters that approximate linear phase in the sense of
maximally flat group delay in the passband. However, we must sacrifice steepness in the
transition region. In addition, acceptable Bessel IIR designs are derived by transforma-
tion only for a relatively limited range of specifications such as sufficiently low cut-off
frequency V
p
.
6.2.3 Frequency Transforms
We have discussed the design of prototype analog lowpass filters with a cut-off fre-
quency V
p
. Although the same procedure can be applied to designing highpass, band-
pass, or bandstop filters, it is much easier to obtain these filters from the desired lowpass
filter using frequency transformations. In addition, most classical filter design tables
only generate lowpass filters and must be converted using spectral transformation into
highpass, bandpass, or bandstop filters. Filter design packages such as MATLAB often
incorporate and perform the frequency transformations directly.
Butterworth highpass filter's transfer function H
hp
(s) can be obtained from the
corresponding lowpass filter's transfer function H(s) by using the relationship
H
hp
(s) = H(s)[
s =
1
s
= H
1
s

: (6:2:12)
For example, consider L = 1. From Table 6.2, we have H(s) = 1=(s ÷1). From (6.2.12),
we obtain
ANALOG FILTERS 253
H
hp
(s) =
1
s ÷1

s =
1
s
=
s
s ÷1
: (6:2:13)
Similarly, we can calculate H
hp
(s) for higher order filters. We can show that the
denominator polynomials of H(s) and H
hp
(s) are the same, but the numerator becomes
s
L
for the Lth-order highpass filters. Thus H
hp
(s) has an additional Lth-order zero at the
origin, and has identical poles s
k
as given in (6.2.10).
Transfer functions of bandpass filters can be obtained from the corresponding low-
pass filters by replacing s with (s
2
÷V
2
m
)=BW. That is,
H
bp
(s) = H(s)

s =
s
2
÷V
2
m
BW
,
(6:2:14)
where V
m
is the center frequency of the bandpass filter and BW is its bandwidth. As
illustrated in Figure 5.3 and defined in (5.1.10) and (5.1.11), the center frequency is
defined as
V
m
=

V
a
V
b

, (6:2:15)
where V
a
and V
b
are the lower and upper cut-off frequencies. The filter bandwidth is
defined by
BW= V
b
÷V
a
: (6:2:16)
Note that for an Lth-order lowpass filter, we obtain a 2Lth-order bandpass filter
transfer function.
For example, consider L = 1. From Table 6.2 and (6.2.14), we have
H
bp
(s) =
1
s ÷1

s =
s
2
÷V
2
m
BW
=
BWs
s
2
÷BWs ÷V
2
m
: (6:2:17)
In general, H
bp
(s) has L zeros at the origin and L pole-pairs.
Bandstop filter transfer functions can be obtained from the corresponding highpass
filters by replacing s in the highpass filter transfer function with (s
2
÷V
2
m
)=BWs. That
is,
H
bs
(s) = H
hp
(s)

s =
s
2
÷V
2
m
BWs
, (6:2:18)
where V
m
is the center frequency defined in (6.2.15) and BW is the bandwidth defined in
(6.2.16).
254 DESIGN AND IMPLEMENTATION OF IIR FILTERS
6.3 Design of IIR Filters
In this section, we discuss the design of digital filters that have an infinite impulse
response. In designing IIR filters, the usual starting point will be an analog filter transfer
function H(s). Because analog filter design is a mature and well-developed field, it is
not surprising that we begin the design of digital IIR filters in the analog domain and
then convert the design into the digital domain. The problem is to determine a digital
filter H(z) which will approximate the performance of the desired analog filter H(s).
There are two methods, the impulse-invariant method and the bilinear transform, for
designing digital IIR filters based on existing analog IIR filters. Instead of designing
the digital IIR filter directly, these methods map the digital filter into an equivalent
analog filter, which can be designed by one of the well-developed analog filter
design methods. The designed analog filter is then mapped back into the desired digital
filter.
The impulse-invariant method preserves the impulse response of the original analog
filter by digitizing the impulse response of analog filter, but not its frequency (magni-
tude) response. Because of inherent aliasing, this method is inappropriate for highpass
or bandstop filters. The bilinear-transform method yields very efficient filters, and is
well suited for the design of frequency selective filters. Digital filters resulting from the
bilinear transform will preserve the magnitude response characteristics of the analog
filters, but not the time domain properties. In general, the impulse-invariant method is
good for simulating analog filters, but the bilinear-transform method is better for
designing frequency selective IIR filters.
6.3.1 Review of IIR Filters
As discussed in Chapters 3 and 4, an IIR filter can be specified by its impulse response
¦h(n), n = 0, 1, . . . , ·¦, I/O difference equation, or transfer function. The general
form of the IIR filter transfer function is defined in (4.3.10) as
H(z) =
¸
L÷1
l=0
b
l
z
÷l
1 ÷
¸
M
m=1
a
m
z
÷m
: (6:3:1)
The design problem is to find the coefficients b
l
and a
m
so that H(z) satisfies the given
specifications. This IIR filter can be realized by the I/O difference equation
y(n) =
¸
L÷1
l=0
b
l
x(n ÷l) ÷
¸
M
m=1
a
m
y(n ÷m): (6:3:2)
The impulse response h(n) of the IIR filter is the output that results when the input is
the unit impulse response defined in (3.1.1). Given the impulse response, the filter
output y(n) can also be obtained by linear convolution expressed as
DESIGN OF IIR FILTERS 255
y(n) = x(n) + h(n) =
¸
·
k=0
h(k)x(n ÷k): (6:3:3)
However, Equation (6.3.3) is not computationally feasible because it uses an infinite
number of coefficients. Therefore we restrict our attention to IIR filters that are
described by the linear difference equation given in (6.3.2).
By factoring the numerator and denominator polynomials of H(z) given in (6.3.1) and
assuming M = L ÷1, the transfer function can be expressed in (4.3.12) as
H(z) = b
0
¸
M
m=1
(z ÷z
m
)
¸
M
m=1
(z ÷p
m
)
, (6:3:4)
where z
m
and p
m
are the mth zero and pole, respectively. For a system to be stable, it is
necessary that all its poles lie strictly inside the unit circle on the z-plane.
6.3.2 Impulse-Invariant Method
The design technique for an impulse-invariant digital filter is illustrated in Figure 6.9.
Assuming the impulse function d(t) is used as a signal source, the output of the analog
filter will be the impulse response h(t). Sampling this continuous-time impulse response
yields the sample values h(nT). In the second signal path, the impulse function d(t) is
sampled first to yield the discrete-time impulse sequence d(n). Filtering this signal by
H(z) yields the impulse response h(n) of the digital filter. If the coefficients of H(z) are
adjusted so that the impulse response coefficients are identical to the previous specified
h(nT), that is,
h(n) = h(nT), (6:3:5)
the digital filter H(z) is the impulse invariant equivalent of the analog filter H(s). An
analog filter H(s) and a digital filter H(z) are impulse invariant if the impulse response of
H(z) is the same as the sampled impulse response of H(s). Thus in effect, we sample the
continuous-time impulse response to produce the discrete-time filter as described by
(6.3.5).
d(t)
h(t)
d(n)
H(z) Sampler
H(s) Sampler
h(nT)
h(n)
Figure 6.9 The concept of impulse-invariant design
256 DESIGN AND IMPLEMENTATION OF IIR FILTERS
The impulse-invariant design is usually not performed directly in the form of (6.3.5).
In practice, the transfer function of an analog filter H(s) is first expanded into a partial-
fraction form
H(s) =
¸
P
i=1
c
i
s ÷s
i
, (6:3:6)
where s = ÷s
i
is the pole of H(s), and c
i
is the residue of the pole at ÷s
i
. Note that we
have assumed there are no multiple poles. Taking the inverse Laplace transform of
(6.3.6) yields
h(t) =
¸
P
i=1
c
i
e
÷s
i
t
, t _ 0, (6:3:7)
which is the impulse response of the analog filter H(s).
The impulse response samples are obtained by setting t equal to nT. From (6.3.5) and
(6.3.7), we have
h(n) =
¸
P
i=1
c
i
e
÷s
i
nT
, n _ 0, (6:3:8)
The z-transform of the sampled impulse response is given by
H(z) =
¸
·
n=0
h(n)z
÷n
=
¸
P
i=1
c
i
¸
·
n=0
(e
÷s
i
T
z
÷1
)
n
=
¸
P
i=1
c
i
1 ÷e
÷s
i
T
z
÷1
: (6:3:9)
The impulse response of H(z) is obtained by taking the inverse z-transform of (6.3.9).
Therefore the filter described in (6.3.9) has an impulse response equivalent to the
sampled impulse response of the analog filter H(s) defined in (6.3.6). Comparing
(6.3.6) with (6.3.9), the parameters of H(z) may be obtained directly from H(s) without
bothering to evaluate h(t) or h(n).
The magnitude response of the digital filter will be scaled by f
s
(= 1=T) due to the
sampling operation. Scaling the magnitude response of the digital filter to approximate
magnitude response of the analog filter requires the multiplication of H(z) by T.
The transfer function of the impulse-invariant digital filter given in (6.3.9) is modified
as
H(z) = T
¸
P
i=1
c
i
1 ÷e
÷s
i
T
z
÷1
: (6:3:10)
The frequency variable ! for the digital filter bears a linear relationship to that for the
analog filter within the operating range of the digital filter. This means that when !
varies from 0 to p around the unit circle in the z-plane, V varies from 0 to p=T along the
jV-axis in the s-plane. Recall that ! = VT as given in (3.1.7). Thus critical frequencies
DESIGN OF IIR FILTERS 257
such as cutoff and bandwidth frequencies specified for the digital filter can be used
directly in the design of the analog filter.
Example 6.6: Consider the analog filter expressed as
H(s) =
0:5(s ÷4)
(s ÷1)(s ÷2)
=
1:5
s ÷1
÷
1
s ÷2
:
The impulse response of the filter is
h(t) = 1:5e
÷t
÷e
÷2t
:
Taking the z-transform and scaling by T yields
H(z) =
1:5T
1 ÷e
÷T
z
÷1
÷
T
1 ÷e
÷2T
z
÷1
:
It is interesting to compare the frequency response of the two filters given in Example
6.6. For the analog filter, the frequency response is
H(V) =
0:5(4 ÷jV)
(1 ÷jV)(2 ÷jV)
:
For the digital filter, we have
H(!) =
1:5T
1 ÷e
÷T
e
÷j!T
÷
T
1 ÷e
÷2T
e
÷j!T
:
The DC response of the analog filter is given by
H(0) = 1, (6:3:11)
and
H(0) =
1:5T
1 ÷e
÷T
÷
T
1 ÷e
÷2T
(6:3:12)
for the digital filter. Thus the responses are different due to aliasing at DC. For a high
sampling rate, T is small and the approximations e
÷T
~ 1 ÷T and e
÷2T
~ 1 ÷2T are
valid. Thus Equation (6.3.12) can be approximated with
H(0) ~
1:5T
1 ÷(1 ÷T)
÷
T
1 ÷(1 ÷2T)
= 1: (6:3:13)
Therefore by using a high sampling rate, the aliasing effect becomes negligible and the
DC gain is one as shown in (6.3.13).
258 DESIGN AND IMPLEMENTATION OF IIR FILTERS
While the impulse-invariant method is straightforward to use, it suffers fromobtaining
a discrete-time system from a continuous-time system by the process of sampling. Recall
that sampling introduces aliasing, and that the frequency response corresponding to the
sequence h(nT) is obtained from (4.4.18) as
H(!) =
1
T
¸
·
k=÷·
H V÷
2pk
T

: (6:3:14)
This is not a one-to-one transformation from the s-plane to the z-plane. Therefore
H(!) =
1
T
H(V) is true only if H(V) = 0 for [V[ _ p=T. As shown in (6.3.14), H(!) is
the aliased version of H(V). Hence the stopband characteristics are maintained ad-
equately if the aliased tails of H(V) are sufficiently small. The passband is also affected,
but this effect is usually less serious. Thus the resulting digital filter does not exactly
meet the original design specifications.
In a bandlimited filter, the magnitude response of the analog filter is negligibly small
at frequencies exceeding half the sampling frequency in order to reduce the aliasing
effect. Thus we must have
[H(!)[ ÷0, for ! _ p: (6:3:15)
This condition can hold for lowpass and bandpass filters, but not for highpass and
bandstop filters.
MATLAB supports the design of impulse invariant digital filters through the func-
tion impinvar in the Signal Processing Toolbox. The s-domain transfer function is first
defined along with the sampling frequency. The function impinvar determines the
numerator and denominator of the z-domain transfer function. The MATLAB com-
mand is expressed as
[bz, az]= impinvar(b, a, Fs)
where bz and az are the numerator and denominator coefficients of a digital filter, Fs is
the sampling rate, and b and a represent coefficients of the analog filter.
6.3.3 Bilinear Transform
As discussed in the previous section, the time-domain impulse-invariant method of filter
design is simple, but has an undesired aliasing effect. This is because the impulse-
invariant method uses the transformation ! = VT or equivalently, z = e
sT
. As dis-
cussed in Section 6.1.3, such mapping leads to aliasing problems. In this section,
we discuss the most commonly used technique for designing IIR filters with pre-
scribed magnitude response specifications ± the bilinear transform. The procedure of
designing digital filters using bilinear transform is illustrated in Figure 6.10. Instead
of designing the digital filter directly, this method maps the digital filter specifications to
an equivalent analog filter, which can be designed by using analog filter design methods
introduced in Section 6.2. The designed analog filter is then mapped back to the desired
digital filter.
DESIGN OF IIR FILTERS 259
Digital filter
specifications
Digital filter
H(z)
Bilinear
transform
Bilinear
transform
w → Ω
w ← Ω
Analog filter
specifications
Analog filter
H(s)
Aanlog filter
design
Figure 6.10 Digital IIR filter design using the bilinear transform
The bilinear transform is a mapping or transformation that relates points on the s-
and z-planes. It is defined as
s =
2
T
z ÷1
z ÷1

=
2
T
1 ÷z
÷1
1 ÷z
÷1

, (6:3:16)
or equivalently,
z =
1 ÷(T=2)s
1 ÷(T=2)s
: (6:3:17)
This is called the bilinear transform because of the linear functions of z in both the
numerator and denominator of (6.3.16).
As discussed in Section 6.1.2, the jV-axis of the s-plane (s = 0) maps onto the unit
circle in the z-plane. The left (s < 0) and right (s > 0) halves of the s-plane map into the
inside and outside of the unit circle, respectively. Because the jV-axis maps onto the unit
circle ([z[ = 1), there is a direct relationship between the s-plane frequency V and the z-
plane frequency !. Substituting s = jV and z = e
j!
into (6.3.16), we have
jV =
2
T
e
j!
÷1
e
j!
÷1

: (6:3:18)
It can be easily shown that the corresponding mapping of frequencies is obtained as
V =
2
T
tan
!
2

, (6:3:19)
or equivalently,
! = 2 tan
÷1
VT
2

: (6:3:20)
Thus the entire jV-axis is compressed into the interval [÷p=T, p=T[ for ! in a one-to-
one manner. The range 0 ÷·portion in the s-plane is mapped onto the 0 ÷p portion
of the unit circle in the z-plane, while the 0 ÷÷·portion in the s-plane is mapped onto
260 DESIGN AND IMPLEMENTATION OF IIR FILTERS
0 1
ΩT
2
p
p
w
−p
Figure 6.11 Plot of transformation given in (6.3.20)
the 0 ÷÷p portion of the unit circle in the z-plane. Each point in the s-plane is uniquely
mapped onto the z-plane. This fundamental relation enables us to locate a point V on
the jV-axis for a given point on the unit circle.
The relationship in (6.3.20) between the frequency variables V and ! is illustrated in
Figure 6.11. The bilinear transform provides a one-to-one mapping of the points along
the jV-axis onto the unit circle, i.e., the entire jV axis is mapped uniquely onto the unit
circle, or onto the Nyquist band [![ _ p. However, the mapping is highly nonlinear. The
point V = 0 is mapped to ! = 0 (or z = 1), and the point V = ·is mapped to ! = p (or
z = ÷1). The entire band VT _ 1 is compressed onto p=2 _ ! _ p. This frequency
compression effect associated with the bilinear transform is known as frequency warp-
ing due to the nonlinearity of the arctangent function given in (6.3.20). This nonlinear
frequency-warping phenomenon must be taken into consideration when designing
digital filters using the bilinear transform. This can be done by pre-warping the critical
frequencies and using frequency scaling.
The bilinear transform guarantees that
H(s)

s=jV
= H(z)

z = e
j!
, (6:3:21)
where H(z) is the transfer function of the digital filter, and H(s) is the transfer function
of an analog filter with the desired frequency characteristics.
6.3.4 Filter Design Using Bilinear Transform
The bilinear transform of an analog filter function H(s) is obtained by simply replacing s
with z using Equation (6.3.16). The filter specifications will be in terms of the critical
frequencies of the digital filter. For example, the critical frequency ! for a lowpass filter
is the bandwidth of the filter, and for a notch filter, it is the notch frequency. If we use
the same critical frequencies for the analog design and then apply the bilinear transform,
the digital filter frequencies would be in error because of the frequency wrapping given
in (6.3.20). Therefore we have to pre-wrap the critical frequencies of the analog filter.
DESIGN OF IIR FILTERS 261
There are three steps involved in the bilinear design procedure. These steps are
summarized as follows:
1. Pre-wrap the critical frequency !
c
of the digital filter using (6.3.19) to obtain the
corresponding analog filter's frequency V
c
.
2. Frequency scale of the designed analog filter H(s) with V
c
to obtain
^
H(s) =
^
H(s)[
s=s=V
c
= H
s
V
c

, (6:3:22)
2. where
^
H(s) is the scaled transfer function corresponding to H(s).
3. Replace s in
^
H(s) by 2(z ÷1)=(z ÷1)T to obtain desired digital filter H(z). That is
H(z) =
^
H(s)[
s=2(z÷1)=(z÷1)T
, (6:3:23)
2. where H(z) is the desired digital filter.
Example 6.7: Consider the transfer function of the simple analog lowpass filter
given as
H(s) =
1
1 ÷s
:
Use this H(s) and the bilinear transform method to design the corresponding
digital lowpass filter whose bandwidth is 1000 Hz and the sampling frequency is
8000 Hz.
The critical frequency for the lowpass filter is the filter bandwidth
!
c
= 2p(1000=8000) radians/sample and T = 1=8000 second.
Step 1:
V
c
=
2
T
tan
!
c
2

=
2
T
tan
2000p
16 000

=
2
T
tan
p
8

=
0:8284
T
:
Step 2: We use frequency scaling to obtain
^
H(s) = H(s)[
s=s=(0:8284=T)
=
0:8284
sT ÷0:8284
:
Step 3: The bilinear transform in (6.3.12) yields the desired transfer function
H(z) =
^
H(s)[
s=2(z÷1)=(z÷1)T
= 0:2929
1 ÷z
÷1
1 ÷0:4142z
÷1
:
262 DESIGN AND IMPLEMENTATION OF IIR FILTERS
MATLAB provides the function bilinear to design digital filters using the bilinear
transform. The transfer function for the analog prototype is first determined. The
numerator and denominator polynomials of the analog prototype are then mapped to
the polynomials for the digital filter using the bilinear transform. For example, the
following MATLAB script can be used for design a lowpass filter using bilinear trans-
form:
Fs = 2000; % Sampling frequency
Wn = 2*pi*500; % Edge frequency
n = 2; % Order of analog filter
[b, a]= butter(n, Wn, `s'); % Design analog filter
[bz, az]= bilinear(b, a, Fs); % Determine digital filter
6.4 Realization of IIR Filters
As discussed earlier, a digital IIR filter can be described by the linear convolution
(6.3.3), the transfer function (6.3.1), or the I/O difference equation (6.3.2). These
equations are equivalent mathematically, but may be different in realization. In DSP
implementation, we have to consider the required operations, memory storage, and the
finite wordlength effects. A given transfer function H(z) can be realized in several forms
or configurations. In this section, we will discuss direct-form I, direct-form II, cascade,
and parallel realizations. Many additional structures such as wave digital filters, ladder
structures, and lattice structures can be found in the reference book [7].
6.4.1 Direct Forms
Given an IIR filter described by (6.3.1), the direct-form I realization is defined by the
I/O Equation (6.3.2). It has L ÷M coefficients and needs L ÷M÷1 memory locations
to store ¦x(n ÷l), l = 0, 1, . . . , L ÷1¦ and ¦y(n ÷m), m = 0, 1, . . . , M¦. It also
requires L ÷M multiplications and L ÷M÷1 additions for implementation on
a DSP system. The detailed signal-flow diagram for L = M÷1 is illustrated in
Figure 4.6.
Example 6.8: Given a second-order IIR filter transfer function
H(z) =
b
0
÷b
1
z
÷1
÷b
2
z
÷2
1 ÷a
1
z
÷1
÷a
2
z
÷2
, (6:4:1)
the I/O difference equation of direct-form I realization is described as
y(n) = b
0
x(n) ÷b
1
x(n ÷1) ÷b
2
x(n ÷2) ÷a
1
y(n ÷1) ÷a
2
y(n ÷2): (6:4:2)
The signal-flow diagram is illustrated in Figure 6.12.
REALIZATION OF IIR FILTERS 263
As shown in Figure 6.12, the IIR filter can be interpreted as the cascade of two
transfer functions H
1
(z) and H
2
(z). That is,
H(z) = H
1
(z)H
2
(z): (6:4:3)
where H
1
(z) = b
0
÷b
1
z
÷1
÷b
2
z
÷2
and H
2
(z) = 1=(1 ÷a
1
z
÷1
÷a
2
z
÷2
). Since multiplica-
tion is commutative, we have
H(z) = H
2
(z)H
1
(z): (6:4:4)
Therefore Figure 6.12 can be redrawn as Figure 6.13.
Note that in Figure 6.13, the intermediate signal w(n) is common to both signal
buffers of H
1
(z) and H
2
(z). There is no need to use two separate buffers, thus these
two signal buffers can be combined into one, shared by both filters as illustrated in
Figure 6.14. We observe that this realization requires three memory locations to realize
the second-order IIR filter, as opposed to six memory locations required for the direct-
form I realization given in Figure 6.12. Therefore the direct-form II realization is called
H
1
(z) H
2
(z)
z
−1
z
−1
z
−1
z
−1
x(n) b
0
b
1
b
2
− a
1
− a
2
y(n)
x(n−1)
x(n−2)
y(n−1)
y(n−2)
Figure 6.12 Direct-form I realization of second-order IIR filter
x(n) y(n)
H
1
(z) H
2
(z)
z
−1
z
−1
z
−1
z
−1
b
0
b
1
b
2
− a
1
− a
2
w(n) w(n)
Figure 6.13 Signal-flow diagram of H(z) = H
2
(z)H
1
(z)
264 DESIGN AND IMPLEMENTATION OF IIR FILTERS
w(n−2)
w(n−1)
x(n) y(n)
b
0
b
2
z
−1
z
−1
− a
2
b
1
− a
1
w(n)
Figure 6.14 Direct-form II realization of second-order IIR filter
the canonical form since it realizes the given transfer function with the smallest possible
numbers of delays, adders, and multipliers.
It is worthwhile verifying that the direct-form II realization does indeed implement
the second-order IIR filter. From Figure 6.14, we have
y(n) = b
0
w(n) ÷b
1
w(n ÷1) ÷b
2
w(n ÷2), (6:4:5)
where
w(n) = x(n) ÷a
1
w(n ÷1) ÷a
2
w(n ÷2): (6:4:6)
Taking the z-transform of both sides of these two equations and re-arranging terms, we
obtain
Y(z) = W(z) b
0
÷b
1
z
÷1
÷b
2
z
÷2

(6:4:7)
and
X(z) = W(z) 1 ÷a
1
z
÷1
÷a
2
z
÷2

: (6:4:8)
The overall transfer function equals to
H(z) =
Y(z)
X(z)
=
b
0
÷b
1
z
÷1
÷b
2
z
÷2
1 ÷a
1
z
÷1
÷a
2
z
÷2
which is identical to (6.4.1). Thus the direct-form II realization described by (6.4.5) and
(6.4.6) is identical to the direct-form I realization described in (6.4.2).
Figure 6.14 can be expanded as Figure 6.15 to realize the general IIR filter defined in
(6.3.1) using the direct-form II structure. The block diagram realization of this system
assumes M = L ÷1. If M = L ÷1, one must draw the maximum number of common
delays. Although direct-form II still satisfies the difference Equation (6.3.2), it does not
implement this difference equation directly. Similar to (6.4.5) and (6.4.6), it is a direct
implementation of a pair of I/O equations:
w(n) = x(n) ÷
¸
M
m=1
a
m
w(n ÷m) (6:4:9)
REALIZATION OF IIR FILTERS 265
b
2
w(n−2)
w(n−1)
b
1
x(n) y(n)
b
0
z
−1
z
−1
− a
2
− a
1
w(n)
− a
M
b
L−1
w(n−L−1)
Figure 6.15 Direct-form II realization of general IIR filter, L = M÷1
and
y(n) =
¸
L÷1
l=0
b
l
w(n ÷l): (6:4:10)
The computed value of w(n) from the first equation is passed into the second equation to
compute the final output y(n).
6.4.2 Cascade Form
The cascade realization of an IIR filter assumes that the transfer function is the product
of first-order and/or second-order IIR sections. By factoring the numerator and the
denominator polynomials of the transfer function H(z) as a product of lower order
polynomials, an IIR filter can be realized as a cascade of low-order filter sections.
Consider the transfer function H(z) given in (6.3.4), it can be expressed as
H(z) = b
0
H
1
(z)H
2
(z) H
K
(z) = b
0
¸
K
k=1
H
k
(z), (6:4:11)
where each H
k
(z) is a first- or second-order IIR filter and K is the total number of
sections. That is
H
k
(z) =
z ÷z
i
z ÷p
j
=
1 ÷b
1k
z
÷1
1 ÷a
1k
z
÷1
, (6:4:12)
or
266 DESIGN AND IMPLEMENTATION OF IIR FILTERS
x(n)
H
1
(z) H
2
(z) H
K
(z)
b
0
y(n)
Figure 6.16 Cascade realization of digital filter
H
k
(z) =
(z ÷z
i
)(z ÷z
j
)
(z ÷p
l
)(z ÷p
m
)
=
1 ÷b
1k
z
÷1
÷b
2k
z
÷2
1 ÷a
1k
z
÷1
÷a
2k
z
÷2
: (6:4:13)
The realization of Equation (6.4.11) in cascade form is illustrated in Figure 6.16. In
this form, any real-valued roots can be left as they are or combined into pairs, and the
complex-conjugate roots must be grouped into the same section to guarantee that all the
coefficients of H
k
(z) are real numbers. Assuming that every H
k
(z) is the second-order
IIR filter described by (6.4.13), the I/O equations describing the time-domain operations
of the cascade realization are expressed as
w
k
(n) = x
k
(n) ÷a
1k
w
k
(n ÷1) ÷a
2k
w
k
(n ÷2), (6:4:14a)
y
k
(n) = w
k
(n) ÷b
1k
w
k
(n ÷1) ÷b
2k
w
k
(n ÷2), (6:4:14b)
x
k÷1
(n) = y
k
(n), (6:4:14c)
for k = 1, 2, . . . , K and
x
1
(n) = b
0
x(n), (6:4:15a)
y(n) = y
k
(n): (6:4:15b)
By different ordering and different pairing, it is possible to obtain many different
cascade realizations for the same transfer function H(z). Ordering means the order of
connecting H
k
(z), and pairing means the grouping of poles and zeros of H(z) to form a
section. These different cascade realizations are mathematically equivalent. In practice,
each cascade realization behaves differently from others due to the finite-wordlength
effects. In DSP hardware implementation, the internal multiplications in each section
will generate a certain amount of roundoff error, which is then propagated into the next
section. The total roundoff noise at the final output will depend on the particular
pairing/ordering. The best ordering is the one that generates the minimum overall
roundoff noise. It is not a simple task to determine the best realization among all
possible cascade realizations. However, the complex-conjugate roots should be paired
together, and we may pair the poles and zeros that are closest to each other in each
section.
In the direct-form realization shown in Figure 6.15, the variation of one parameter
will affect the locations of all the poles of H(z). In a cascade realization, the variation of
one parameter will affect only pole(s) in that section. Therefore the cascade realization is
less sensitive to parameter variation (due to coefficient quantization, etc.) than the
direct-form structure. In practical implementations of digital IIR filters, the cascade
form is preferred.
REALIZATION OF IIR FILTERS 267
Example 6.9: Given the second-order IIR filter
H(z) =
0:5(z
2
÷0:36)
z
2
÷0:1z ÷0:72
,
realize it using cascade form in terms of first-order sections.
By factoring the numerator and denominator polynomials of H(z), we obtain
H(z) =
0:5(1 ÷0:6z
÷1
)(1 ÷0:6z
÷1
)
(1 ÷0:9z
÷1
)(1 ÷0:8z
÷1
)
:
By different pairings of poles and zeros, there are four different realizations of
H(z). For example, we choose
H
1
(z) =
1 ÷0:6z
÷1
1 ÷0:9z
÷1
and H
2
(z) =
1 ÷0:6z
÷1
1 ÷0:8z
÷1
:
The IIR filter can be realized by the cascade form expressed as
H(z) = 0:5H
1
(z)H
2
(z):
6.4.3 Parallel Form
The expression of H(z) in a partial-fraction expansion leads to another canonical
structure called the parallel form. It is expressed as
H(z) = c ÷H
1
(z) ÷H
2
(z) ÷ ÷H
k
(z), (6:4:16)
where c is a constant, K is a positive integer, and H
k
(z) are transfer functions of first- or
second-order IIR filters with real coefficients. That is,
H
k
(z) =
b
0k
1 ÷a
1k
z
÷1
, (6:4:17)
or
H
k
(z) =
b
0k
÷b
1k
z
÷1
1 ÷a
1k
z
÷1
÷a
2k
z
÷2
: (6:4:18)
The realization of Equation (6.4.16) in parallel form is illustrated in Figure 6.17. In
order to produce real-valued coefficients in the filter structure, the terms in the partial-
fraction-expansion corresponding to complex-conjugate pole pairs must be combined
into second-order terms. Each second-order section can be implemented as direct-form
II as shown in Figure 6.14, or direct-form I as shown in Figure 6.12.
268 DESIGN AND IMPLEMENTATION OF IIR FILTERS
c
0
H
2
(z)
H
1
(z)
H
K
(z)
x(n) y(n)
Figure 6.17 A parallel realization of digital IIR filter
The variation of parameters in a parallel form affects only the poles of the H
k
(z)
associated with the parameters. The variation of any parameter in the direct-form
realization will affect all the poles of H(z). Therefore the pole sensitivity of a parallel
realization is less than that of the direct form.
Example 6.10: Consider the transfer function H(z) as given in Example 6.9, we can
express
H
/
(z) =
H(z)
z
=
0:5 1 ÷0:6z
÷1

1 ÷0:6z
÷1

z 1 ÷0:9z
÷1
( ) 1 ÷0:8z
÷1
( )
=
A
z
÷
B
z ÷0:9
÷
C
z ÷0:8
,
where
A = zH
/
(z)[
z=0
= 0:25,
B = (z ÷0:9)H
/
(z)[
z=÷0:9
= 0:147, and
C = (z ÷0:8)H
/
(z)[
z=0:8
= 0:103:
We can obtain
H(z) = 0:25 ÷
0:147
1 ÷0:9z
÷1
÷
0:103
1 ÷0:8z
÷1
:
6.4.4 Realization Using MATLAB
The cascade realization of an IIR transfer function H(z) involves its factorization in the
form of (6.3.4). This can be done in MATLAB using the function roots. For example,
the statement
r = roots(b);
will return the roots of the numerator vector b containing the coefficients of polynomial
in z
÷1
in ascending power of z
÷1
in the output vector r. Similarly, we can use
REALIZATION OF IIR FILTERS 269
d = roots(a);
to obtain the roots of the denominator vector a in the output vector d. From the com-
puted roots, the coefficients of each section can be determined by pole±zero pairings.
A much simpler approach is to use the function tf2zp in the Signal Processing
Toolbox, which finds the zeros, poles, and gains of systems in transfer functions of
single-input or multiple-output form. For example, the statement
[z, p, c]= tf2zp(b, a);
will return the zero locations in the columns of matrix z, the pole locations in the
column vector p, and the gains for each numerator transfer function in vector c.
Vector a specifies the coefficients of the denominator in descending powers of z
÷1
,
and the matrix b indicates the numerator coefficients with as many rows as there are
outputs.
Example 6.11: The zeros, poles, and gain of the system
H(z) =
2z
÷1
÷3z
÷2
1 ÷0:4z
÷1
÷z
÷2
can be obtained using the MATLAB script as follows:
b = [2, 3];
a = [1, 0.4, 1];
[z, p, c]= tf2zp(b,a);
MATLAB also provides a useful function zp2sos in the Signal Processing Toolbox
to convert a zero-pole-gain representation of a given system to an equivalent represen-
tation of second-order sections. The function
[sos, G]= zp2sos(z, p, c);
finds the overall gain G and a matrix sos containing the coefficients of each second-
order section of the equivalent transfer function H(z) determined from its zero±pole
form. The zeros and poles must be real or in complex-conjugate pairs. The matrix sos is
a K ×6 matrix
sos =
b
01
b
11
b
21
a
01
a
11
a
21
b
02
b
12
b
22
a
02
a
12
a
22
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
b
0K
b
1K
b
2K
a
0K
a
1K
a
2K

¸
¸
¸
¸
¸
, (6:4:19)
whose rows contain the numerator and denominator coefficients, b
ik
and a
ik
, i = 0, 1, 2
of the kth second-order section H
k
(z). The overall transfer function is expressed as
H(z) =
¸
K
k=1
H
k
(z) =
¸
K
k=1
b
0k
÷b
1k
z
÷1
÷b
2k
z
÷2
a
0k
÷a
1k
z
÷1
÷a
2k
z
÷2
: (6:4:20)
270 DESIGN AND IMPLEMENTATION OF IIR FILTERS
The parallel realizations discussed in Section 6.4.3 can be developed in MATLAB
using the function residuez in the Signal Processing Toolbox. This function converts
the transfer function expressed as (6.3.1) to the partial-fraction-expansion (or residue)
form as (6.4.16). The function
[r, p, c]= residuez(b, a);
returns that the column vector r contains the residues, p contains the pole locations,
and c contains the direct terms.
6.5 Design of IIR Filters Using MATLAB
As discussed in Chapter 5, digital filter design is a process of determining the values of
the filter coefficients for given specifications. This is not an easy task and is generally
better to be performed using computer software packages. A filter design package can
be used to evaluate the filter design methods, to calculate the filter coefficients, and to
simulate the filter's magnitude and phase responses. MATLAB is capable of designing
Butterworth, Chebyshev I, Chebyshev II, and elliptic IIR filters in four different types of
filters: lowpass, highpass, bandpass, and bandstop.
The Signal Processing Toolbox provides a variety of M-files for designing IIR filters.
The IIR filter design using MATLAB requires two processes. First, the filter order N
and the frequency-scaling factor Wn are determined from the given specifications. The
coefficients of the filter are then determined using these two parameters. In the first step,
the MATLAB functions to be used for estimating filter order are
[N, Wn]= buttord(Wp, Ws, Rp, Rs);
[N, Wn]= cheb1ord(Wp, Ws, Rp, Rs);
[N, Wn]= cheb2ord(Wp, Ws, Rp, Rs);
[N, Wn]= ellip(Wp, Ws, Rp, Rs);
for Butterworth, Chebyshev type I, Chebyshev type II, and elliptic filters, respectively.
The parameters Wp and Ws are the normalized passband and stopband edge frequencies,
respectively. The range of Wp and Ws are between 0 and 1, where 1 corresponds to the
Nyquist frequency ( f
N
= f
s
=2). The parameters Rp and Rs are the passband ripple and
the minimum stopband attenuation specified in dB, respectively. These four functions
return the order N and the frequency scaling factor Wn. These two parameters are needed
in the second step of IIR filter design using MATLAB.
For lowpass filters, the normalized frequency range of passband is 0 < F < Wp, the
stopband is Ws < F < 1, and Wp < Ws. For highpass filters, the normalized frequency
range of stopband is 0 < F < Ws, the passband is Wp < F < 1, and Wp > Ws. For
bandpass and bandstop filters, Wp and Ws are two-element vectors that specify the
transition bandages, with the lower-frequency edge being the first element of the vector,
and N is half of the order of the filter to be designed.
In the second step of designing IIR filters based on the bilinear transformation, the
Signal Processing Toolbox provides the following functions:
[b, a]= butter(N, Wn);
[b, a]= cheby1(N, Rp, Wn);
DESIGN OF IIR FILTERS USING MATLAB 271
[b, a]= cheby2(N, Rs, Wn);
[b, a]= ellip(N, Rp, Rs, Wn);
The input parameters N and Wn are determined in the order estimation stage. These
functions return the filter coefficients in length N÷1 row vectors b and a with coeffi-
cients in descending powers of z
÷1
. The form of the transfer function obtained is given
by
H(z) =
b(1) ÷b(2)z
÷1
÷ ÷b(N ÷1)z
÷N
1 ÷a(2)z
÷1
÷ ÷a(N ÷1)z
÷N
: (6:5:1)
As introduced in Chapter 4, the frequency response of digital filters can be computed
using the function freqz, which returns the frequency response H(!) of a given
numerator and denominator coefficients in vectors b and a, respectively.
Example 6.12: Design a lowpass Butterworth filter with less than 1.0 dB of ripple
from 0 to 800 Hz, and at least 20 dB of stopband attenuation from 1600 Hz to the
Nyquist frequency 4000 Hz.
The MATLAB script (exam6_12.m in the software package) for designing the
specified filter is listed as follows:
Wp = 800/4000; Ws = 1600/4000;
Rp = 1.0; Rs = 20.0;
[N, Wn]= buttord(Wp, Ws, Rp, Rs);
[b, a]= butter(N, Wn);
freqz(b, a, 512, 8000);
The Butterworth filter coefficients are returned via vectors b and a by MATLAB
function butter(N,Wn). The magnitude and phase responses of the designed
fourth-order IIR filter are shown in Figure 6.18. This filter will be used for the IIR
filter experiments in Sections 6.7.
Example 6.13: Design a bandpass filter with passband of 100 Hz to 200 Hz and the
sampling rate is 1 kHz. The passband ripple is less than 3 dB and the stopband
attenuation is at least 30 dB by 50 Hz out on both sides of the passband.
The MATLAB script (exam6_13.m in the software package) for designing the
specified bandpass filter is listed as follows:
Wp = [100 200]/500; Ws = [50 250]/500;
Rp = 3; Rs = 30;
[N, Wn]= buttord(Wp, Ws, Rp, Rs);
[b, a]= butter(N, Wn);
freqz(b, a, 128, 1000);
The magnitude and phase responses of the designed bandpass filter are shown in
Figure 6.19.
272 DESIGN AND IMPLEMENTATION OF IIR FILTERS
0 500 1000 1500 2000 2500 3000 3500 4000
−400
−300
−200
−100
0
Frequency (Hz)
P
h
a
s
e

(
d
e
g
r
e
e
s
)
0 500 1000 1500 2000 2500 3000 3500 4000
−250
−200
−150
−100
−50
0
Frequency (Hz)
M
a
g
n
i
t
u
d
e

(
d
B
)
Figure 6.18 Frequency response of fourth-order Butterworth lowpass filter
0
100
0
−100
−200
−300
50 100 150 200 250
Frequency (Hz)
M
a
g
n
i
t
u
d
e

(
d
B
)
300 350 400 450 500
0
200
0
−600
−800
−1000
50 100 150 200 250
Frequency (Hz)
P
h
a
s
e

(
d
e
g
r
e
e
s
)
300 350 400 450 500
−200
−400
Figure 6.19 Frequency response of the designed bandpass filter
6.6 Implementation Considerations
As discussed in Section 6.4, the common IIR filter structures are direct, parallel, and
cascade forms. The cascade forms are most often employed in practical applications for
the reasons concerning quantization effects and DSP implementation. These issues will
be discussed in this section.
IMPLEMENTATION CONSIDERATIONS 273
6.6.1 Stability
The IIR filter described by the transfer function given in (6.3.4) is stable if all the poles
lie within the unit circle. That is,
[p
m
[ < 1, m = 1, 2, . . . , M: (6:6:1)
In this case, we can show that
lim
n÷·
h(n) = 0: (6:6:2)
If [p
m
[ > 1 for any 0 _ m _ M, then the IIR filter defined in (6.3.4) is unstable since
lim
n÷·
h(n) ÷·: (6:6:3)
In addition, an IIR filter is unstable if H(z) has multiple-order pole(s) on the unit circle.
For example, if H(z) = z=(z ÷1)
2
, there is a second-order pole at z = 1. The impulse
response of the system is h(n) = n, which is an unstable system as defined in (6.6.3).
Example 6.14: Given the transfer function
H(z) =
1
1 ÷az
÷1
,
the impulse response of the system is
h(n) = a
n
, n _ 0:
If the pole is inside the unit circle, that is, [a[ < 1, the impulse response
lim
n÷·
h(n) = lim
n÷·
a
n
= 0:
Thus the IIR filter is stable. However,
lim
n÷·
h(n) = lim
n÷·
a
n
÷· if [a[ > 1:
Thus the IIR filter is unstable for [a[ > 1.
An IIR filter is marginally stable (or oscillatory bounded) if
lim
n÷·
h(n) = c, (6:6:4)
where c is a non-zero constant. For example, H(z) = 1=1 ÷z
÷1
. There is a first-order
pole on the unit circle. It is easy to show that the impulse response oscillates between ±1
since h(n) = (÷1)
n
, n _ 0.
274 DESIGN AND IMPLEMENTATION OF IIR FILTERS
Consider the second-order IIR filter defined by equation (6.4.1). The denominator
can be factored as
1 ÷a
1
z
÷1
÷a
2
z
÷2
= (1 ÷p
1
z
÷1
)(1 ÷p
2
z
÷1
), (6:6:5)
where
a
1
= ÷(p
1
÷p
2
) (6:6:6)
and
a
2
= p
1
p
2
: (6:6:7)
The poles must lie inside the unit circle for stability, that is [p
1
[ < 1 and [p
2
[ < 1. From
(6.6.7), we obtain
[a
2
[ = [p
1
p
2
[ < 1: (6:6:8)
The corresponding condition on a
1
can be derived from the Schur±Cohn stability test
and is given by
[a
1
[ < 1 ÷a
2
: (6:6:9)
Stability conditions (6.6.8) and (6.6.9) are illustrated in Figure 6.20, which shows the
resulting stability triangle in the a
1
÷a
2
plane. That is, the second-order IIR filter is
stable if and only if the coefficients define a point (a
1
, a
2
) that lies inside the stability
triangle.
6.6.2 Finite-Precision Effects and Solutions
As discussed in Chapter 3, there are four types of quantization effects in digital filters ±
input quantization, coefficient quantization, roundoff errors, and overflow. In practice,
the digital filter coefficients obtained from a filter design package are quantized to a
finite number of bits so that the filter can be implemented using DSP hardware. The
filter coefficients, b
l
and a
m
, of the discrete-time filter defined by (6.3.1) and (6.3.2) are
determined by the filter design techniques introduced in Section 6.3, or by a filter design
1
a
2
a
1
−1
2 −2
a
2
=1
− a
1
= 1 + a
2
a
1
= 1 + a
2
Figure 6.20 Region of coefficient values for a stable second-order IIR filter
IMPLEMENTATION CONSIDERATIONS 275
package such as MATLAB that uses double-precision floating-point format to repre-
sent filter coefficients. Let b
/
l
and a
/
m
denote the quantized values corresponding to b
l
and a
m
, respectively. The I/O equation that can be actually implemented is given by
Equation (3.5.10), and its transfer function is expressed as
H
/
(z) =
¸
L÷1
l=0
b
/
l
z
÷l
1 ÷
¸
M
m=1
a
/
m
z
÷m
: (6:6:10)
Similar to the concept of input quantization discussed in Section 3.5, the nonlinear
operation of coefficient quantization can be modeled as a linear process that introduces
a quantization noise expressed as
b
/
l
= Q[b
l
[ = b
l
÷e(l) (6:6:11)
and
a
/
m
= Q[a
m
[ = a
m
÷e(m), (6:6:12)
where the coefficient quantization errors e(l) and e(m) can be assumed to be a random
noise that has zero mean and variance as defined in (3.5.6).
If the wordlength is not large enough, some undesirable effects occur. For ex-
ample, the frequency characteristics such as magnitude and phase responses of H
/
(z)
may be different from those of H(z). In addition, for high-order filters whose poles
are closely clustered in the z-plane, small changes in the denominator coefficients can
cause large shifts in the location of the poles. If the poles of H(z) are close to the
unit circle, the pole(s) of H
/
(z) may move outside the unit circle after coefficient
quantization, resulting in an unstable implementation. These undesired effects are
more serious when higher-order filters are implemented using the direct-form I and II
realizations discussed in Section 6.4. Therefore the cascade and parallel realizations are
preferred in practical DSP implementations with each H
k
(z) in a first- or second-order
section.
Example 6.15: Given the IIR filter with transfer function
H(z) =
1
1 ÷0:9z
÷1
÷0:2z
÷2
,
the poles are located at z = 0:4 and z = 0:5. This filter can be realized in the
cascade form as
H(z) = H
1
(z)H
2
(z),
where H
1
(z) = 1=(1 ÷0:4z
÷1
) and H
2
(z) = 1=(1 ÷0:5z
÷1
):
Assuming that this IIR filter is implemented in a 4-bit (a sign bit plus 3 data
bits) DSP hardware, 0.9, 0.2, 0.4, and 0.5 are quantized to 0.875, 0.125, 0.375, and
0.5, respectively. Therefore the direct-form realization is described as
276 DESIGN AND IMPLEMENTATION OF IIR FILTERS
H
/
(z) =
1
1 ÷0:875z
÷1
÷0:125z
÷2
and the cascade realization is expressed as
H
//
(z) =
1
1 ÷0:375z
÷1

1
1 ÷0:5z
÷1
:
The pole locations of the direct-form H
/
(z) are z = 0:18 and z = 0:695, and the
pole locations of the cascade form H
//
(z) are z = 0:375 and z = 0:5. Therefore the
poles of cascade realization are closer to the desired H(z).
In practice, one must always check the stability of the filter with the quantized
coefficients. The problem of coefficient quantization may be studied by examining
pole locations in the z-plane. For a second-order IIR filter given in (6.4.1), we can
place the poles near z = ±1 with much less accuracy than elsewhere in the z-plane. Since
the second-order IIR filters are the building blocks of the cascade and parallel forms, we
can conclude that narrowband lowpass (or highpass) filters will be most sensitive to
coefficient quantization because their poles close to z = 1 or (z = ÷1). In summary, the
cascade form is recommended for the implementation of high-order narrowband IIR
filters that have closely clustered poles.
As discussed in Chapter 3, the effect of the input quantization noise on the output can
be computed as
s
2
y;e
=
s
2
e
2pj

z
÷1
H(z)H(z
÷1
)dz, (6:6:13)
where s
2
e
= 2
÷2B
=3 is defined by (3.5.6). The integration around the unit circle [z[ = 1 in
the counterclockwise direction can be evaluated using the residue method introduced in
Chapter 4 for the inverse z-transform.
Example 6.16: Consider an IIR filter expressed as
H(z) =
1
1 ÷az
÷1
, [a[ < 1,
and the input signal x(n) is an 8-bit data. The noise power due to input quantiza-
tion is s
2
e
= 2
÷16
=3. Since
R
z=a
= (z ÷a)
1
(z ÷a)(1 ÷az)

z=a
=
1
1 ÷a
2
,
the noise power at the output of the filter is calculated as
s
2
y;e
=
2
÷16
3(1 ÷a
2
)
:
IMPLEMENTATION CONSIDERATIONS 277
As shown in (3.5.11), the rounding of 2B-bit product to B bits introduces the roundoff
noise, which has zero-mean and its power is defined by (3.5.6). Roundoff errors can be
trapped into the feedback loops of IIR filters and can be amplified. In the cascade
realization, the output noise power due to the roundoff noise produced at the previ-
ous section may be evaluated using (6.6.13). Therefore the order in which individual
sections are cascaded also influences the output noise power due to roundoff.
Most modern DSP chips (such as the TMS320C55x) solve this problem by using
double-precision accumulator(s) with additional guard bits that can perform many
multiplication±accumulation operations without roundoff errors before the final result
in the accumulator is rounded.
As discussed in Section 3.6, when digital filters are implemented using finite word-
length, we try to optimize the ratio of signal power to the power of the quantization
noise. This involves a trade-off with the probability of arithmetic overflow. The most
effective technique in preventing overflow of intermediate results in filter computation is
by introducing appropriate scaling factors at various nodes within the filter stages. The
optimization is achieved by introducing scaling factors to keep the signal level as high as
possible without getting overflow. For IIR filters, since the previous output is fed back,
arithmetic overflow in computing an output value can be a serious problem. A detailed
analysis related to scaling is available in a reference text [12].
Example 6.17: Consider the first-order IIR filter with scaling factor a described by
H(z) =
a
1 ÷az
÷1
,
where stability requires that [a[ < 1. The actual implementation of this filter is
illustrated in Figure 6.21. The goal of including the scaling factor a is to ensure
that the values of y(n) will not exceed 1 in magnitude. Suppose that x(n) is a
sinusoidal signal of frequency !
0
, then the amplitude of the output is a factor of
[H(!
0
)[. For such signals, the gain of H(z) is
max
!
[H(!)[ =
a
1 ÷[a[
:
Thus if the signals being considered are sinusoidal, a suitable scaling factor is
given by
a < 1 ÷[a[:
a
Q
z
−1
a
y(n)
x(n)
Figure 6.21 Implementation of a first-order IIR section with scaling factor
278 DESIGN AND IMPLEMENTATION OF IIR FILTERS
6.6.3 Software Implementations
As discussed in Chapter 1, the implementation of a digital filtering algorithm is often
carried out on a general-purpose computer to verify that the designed filter indeed meets
the goals of the application. In addition, such a software implementation may be
adequate if the application does not require real-time processing.
For computer software implementation, we describe the filter in the formof a set of I/O
difference equations. For example, a direct-form II realization of IIR filter is defined by
(6.3.2). The MATLAB function filter in the Signal Processing Toolbox implements
the IIR filter. The basic forms of this function are
y = filter(b, a, x);
y = filter(b, a, x, zi);
The numerator and denominator coefficients are contained in the vectors b and a
respectively. The first element of vector a, a(1), has been assumed to be equal to 1.
The input vector is x and the output vector generated by the filter is y. At the beginning,
the initial conditions (data in the signal buffer) are set to zero. However, they can be
specified in the vector zi to reduce transients.
The direct-form realization of IIR filters can be implemented using following C
function (iir.c in the software package):
/****************************************************************
* IIR.C ± This function performs IIR filtering *
* *
* na÷1 nb÷1 *
* yn = sum ai*x(n÷i) ÷ sum bj*y(n÷j) *
* i=0 j=1 *
* *
****************************************************************/
float iir(float *x, int na, float *a, float *y, int nb, float *b,
int maxa, int maxb)
{
float yn; /* Output of IIR filter */
float yn1, yn2; /* Temporary storage */
int i, j; /* Indexes */
yn1 = (float) 0.0; /* y1(n)= 0. */
yn2 = (float) 0.0; /* y2(n)= 0. */
for(i = 0; i <= na÷1; ++i)
{
yn1 = yn1 ÷ x[maxa÷1÷i]*a[i];
/* FIR filtering of x(n)to get y1(n) */
}
IMPLEMENTATION CONSIDERATIONS 279
for(j = 1; j <= nb÷1; ++j)
{
yn2 = yn2 ÷ y[maxb÷j]*b[j];
/* FIR filtering of y(n)to get y2(n) */
}
yn = yn1÷yn2; /* y(n) = y1(n)÷y2(n) */
return(yn); /* Return y(n)to the main function */
}
6.6.4 Practical Applications
Consider a simple second-order resonator filter whose frequency response is dominated
by a single peak at frequency !
0
. To make a peak at ! = !
0
, we place a pair of complex-
conjugate poles at
p
i
= r
p
e
±j!
0
, (6:6:14)
where 0 < r
p
< 1. The transfer function can be expressed as
H(z) =
A
(1 ÷r
p
e
j!
0
z
÷1
)(1 ÷r
p
e
÷j!
0
z
÷1
)
=
A
1 ÷2r
p
cos(!
0
)z
÷1
÷r
2
p
z
÷2
=
A
1 ÷a
1
z
÷1
÷a
2
z
÷2
,
(6:6:15)
where A is a fixed gain used to normalize the filter to unity at !
0
. That is, [H(!
0
)[ = 1.
The direct-form realization is shown in Figure 6.22.
The magnitude response of this normalized filter is given by
[H(!
0
)[
z=e
÷j!
0
=
A
1 ÷r
p
e
j!
0
e
÷j!
0

1 ÷r
p
e
÷j!
0
e
÷j!
0

= 1: (6:6:16)
This condition can be solved to obtain the gain
A = (1 ÷r
p

1 ÷r
p
e
÷2j!
0

= 1 ÷r
p

1 ÷2r
p
cos(2!
0
) ÷r
2
p

: (6:6:17)
x(n)
A
2r
p
cosw
0
y(n)
−r
p
z
−1
z
−1
2
Figure 6.22 Signal flow graph of second-order resonator filter
280 DESIGN AND IMPLEMENTATION OF IIR FILTERS
The 3-dB bandwidth of the filter is defined as
20 log
10
H(!)
H(!
0
)

= 10 log
10
1
2

= ÷3 dB: (6:6:18)
This equation is equivalent to
[H(!)[
2
=
1
2
[H(!
0
)[
2
=
1
2
: (6:6:19)
There are two solutions on both sides of !
0
, and the bandwidth is the difference between
these two frequencies. When the poles are close to the unit circle, the BW is approxi-
mated as
BW%2 1 ÷r
p

: (6:6:20)
This design criterion determines the value of r
p
for a given BW. The closer r
p
is to one,
the sharper the peak, and the longer it takes for the filter to reach its steady-state
response. From (6.6.15), the I/O difference equation of resonator is given by
y(n) = Ax(n) ÷a
1
y(n ÷1) ÷a
2
y(n ÷2), (6:6:21)
where
a
1
= ÷2r
p
cos !
0
(6:6:22a)
and
a
2
= r
2
p
: (6:6:22b)
A recursive oscillator is a very useful tool for generating sinusoidal waveforms. The
method is to use a marginally stable two-pole resonator where the complex-conjugate
poles lie on the unit circle (r
p
= 1). This recursive oscillator is the most efficient way for
generating a sinusoidal waveform, particularly if the quadrature signals (sine and cosine
signals) are required.
Consider two causal impulse responses
h
c
(n) = cos(!
0
n)u(n) (6:6:23a)
and
h
s
(n) = sin(!
0
n)u(n), (6:6:23b)
where u(n) is the unit step function. The corresponding system transfer functions are
H
c
(z) =
1 ÷cos(!
0
)z
÷1
1 ÷2 cos(!
0
)z
÷1
÷z
÷2
(6:6:24a)
IMPLEMENTATION CONSIDERATIONS 281
and
H
s
(z) =
sin(!
0
)z
÷1
1 ÷2 cos(!
0
)z
÷1
÷z
÷2
: (6:6:24b)
A two-output recursive structure with these system transfer functions is illustrated in
Figure 6.23. The implementation requires just two data memory locations and two
multiplications per sample. The output equations are
y
c
(n) = w(n) ÷cos(!
0
)w(n ÷1) (6:6:25a)
and
y
s
(n) = sin(!
0
)w(n ÷1), (6:6:25b)
where w(n) is an internal state variable that is updated as
w(n) = 2 cos(!
0
)w(n ÷1) ÷w(n ÷2): (6:6:26)
An impulse signal Ad(n) is applied to excite the oscillator, which is equivalent to
presetting w(n) and w(n ÷1) to the following initial conditions:
w(0) = A (6:6:27a)
and
w(÷1) = 0: (6:6:27b)
The waveform accuracy is limited primarily by the DSP processor wordlength. For
example, quantization of the coefficient cos(!
0
) causes the actual output frequency to
differ slightly from the ideal frequency !
0
.
For some applications, only a sinewave is required. From equations (6.6.21), (6.2.22a)
and (6.6.22b) using the conditions that x(n) = Ad(n) and r
p
= 1, we can obtain the
sinusoidal function
sin(w
0
) cos(w
0
) 2
+

w(n−2)
w(n−1)
w(n)
+

y
c
(n)
y
s
(n)
z
−1
z

1
Figure 6.23 Recursive quadrature oscillator
282 DESIGN AND IMPLEMENTATION OF IIR FILTERS
y
s
(n) = Ax(n) ÷a
1
y
s
(n ÷1) ÷a
2
y
s
(n ÷2)
= 2 cos(!
0
)y
s
(n ÷1) ÷y
s
(n ÷2) (6:6:28)
with the initial conditions
y
s
(1) = Asin(!
0
) (6:6:29a)
and
y
s
(0) = 0: (6:6:29b)
The oscillating frequency defined by Equation (6.6.28) is determined from its coefficient
a
1
and its sampling frequency f
s
, expressed as
f = cos
÷1
[a
1
[
2

f
s
2p
Hz, (6:6:30)
where the coefficient [a
1
[ _ 2.
The sinewave generator using resonator can be realized from the recursive computa-
tion given in (6.6.28). The implementation using the TMS320C55x assembly language is
listed as follows:
mov cos_w,T1
mpym *AR1÷,T1,AC0 ; AC0 = cos(w)*y[n÷1]
sub *AR1-<#16,AC0,AC1 ; AC1 = cos(w)*y[n÷1]÷ y[n÷2]
add AC0,AC1 ; AC1 = 2*cos(w)*y[n÷1]÷ y[n÷2]
[ [ delay *AR1 ; y[n÷2]= y[n÷1]
mov rnd(hi(AC1)),*AR1 ; y[n÷1]= y[n]
mov rnd(hi(AC1)),*AR0÷ ; y[n]= 2*cos(w)*y[n÷1]÷ y[n÷2]
In the program, AR1 is the pointer for the signal buffer. The output sinewave samples
are stored in the output buffer pointed by AR0. Due to the limited wordlength, the
quantization error of fixed-point DSPs such as the TMSC320C55x could be severe for
the recursive computation.
A simple parametric equalizer filter can be designed from a resonator given in (6.6.15)
by adding a pair of zeros near the poles at the same angles as the poles. That is, placing
the complex-conjugate poles at
z
i
= r
z
e
±j!
0
, (6:6:31)
where 0 < r
z
< 1. Thus the transfer function given in (6.6.15) becomes
H(z) =
(1 ÷r
z
e
j!
0
z
÷1
)(1 ÷r
z
e
÷j!
0
z
÷1
)
(1 ÷r
p
e
j!
0
z
÷1
)(1 ÷r
p
e
÷j!
0
z
÷1
)
=
1 ÷2r
z
cos(!
0
)z
÷1
÷r
2
z
z
÷2
1 ÷2r
p
cos(!
0
)z
÷1
÷r
2
p
z
÷2
=
1 ÷b
1
z
÷1
÷b
2
z
÷2
1 ÷a
1
z
÷1
÷a
2
z
÷2
:
(6:6:32)
IMPLEMENTATION CONSIDERATIONS 283
When r
z
< r
p
, the pole dominates over the zero because it is closer to the unit circle
than the zero does. Thus it generates a peak in the frequency response at ! = !
0
. When
r
z
> r
p
, the zero dominates over the pole, thus providing a dip in the frequency
response. When the pole and zero are very close to each other, the effects of the poles
and zeros are reduced, resulting in a flat response. Therefore Equation (6.6.32) provides
a boost if r
z
< r
p
, or a reduction if r
z
> r
p
. The amount of gain and attenuation is
controlled by the difference between r
p
and r
z
. The distance from r
p
to the unit circle will
determine the bandwidth of the equalizer.
6.7 Software Developments and Experiments Using the
TMS320C55x
The digital IIR filters are widely used for practical DSP applications. In the previous
sections, we discussed the characteristics, design, realization, and implementation of IIR
filters. The experiments given in this section demonstrate DSP system design process
using an IIR filter as example. We will also discuss some practical considerations for
real-time applications.
As shown in Figure 1.8, a DSP system design usually consists of several steps, such as
system requirements and specifications, algorithm development and simulation, soft-
ware development and debugging, as well as system integration and testing. In this
section, we will use an IIR filter as an example to show these steps with an emphasis on
software development. First, we define the filter specifications such as the filter type,
passband and stopband frequency ranges, passband ripple, and stopband attenuation.
We then use MATLAB to design the filter and simulate its performance. After the
simulation results meet the given specifications, we begin the software development
process. We start with writing a C program with the floating-point data format in order
to compare with MATLAB simulation results. We then measure the filter performance
and improve its efficiency by using fixed-point C implementation and C55x assembly
language. Finally, the design is integrated into the DSP system and is tested again.
Figure 6.24 shows a commonly used flow chart of DSP software development. In the
past, software development was heavily concentrated in stage 3, while stage 2 was
skipped. With the rapid improvement of DSP compiler technologies in recent years, C
compilers have been widely used throughout stages 1 and 2 of the design process. Aided
by compiler optimization features such as intrinsic as well as fast DSP processor speed,
real-time DSP applications are widely implemented using the mixed C and assembly
code. In the first experiment, we will use the floating-point C code to implement an
IIR filter in the first stage as shown in Figure 6.24. Developing code in stage 1 does not
require knowledge of the DSP processors and is suitable for algorithm development and
analysis. The second and third experiments emphasize the use of C compiler optimiza-
tion, data type management, and intrinsic for stage 2 of the design process. The third
stage requires the longest development time because assembly language programming
is much more difficult than C language programming. The last experiment uses
the assembly code in order to compare it with previous experiments. In general, the
assembly code is proven to be the most efficient in implementing DSP algorithms such
as filtering that require intensive multiply/accumulate operations, while C code can do
well in data manipulation such as data formatting and arrangement.
284 DESIGN AND IMPLEMENTATION OF IIR FILTERS
Stage 1
Stage 2
Stage 3
no
no
no
no
no
yes
yes
yes
yes
yes
DSP Algorithm
Floating-point C
Efficient?
Fixed-point C
Efficient?
Refine?
Assembly code
Efficient?
Refine?
Complete
Figure 6.24 Flow chart for DSP software development
6.7.1 Design of IIR Filter
Digital filter coefficients can be determined by filter design software packages such as
MATLAB for given specifications. As mentioned in Section 6.4, high-order IIR filters
are often implemented in the form of cascade or parallel second-order sections for real-
time applications. For instance, the fourth-order Butterworth filter given by Example
6.12 can be realized in the cascade direct-form II structure. The following MATLAB
script (s671.m in the software package) shows a lowpass IIR Butterworth filter design
process.
%
% Filter specifications
%
Fs = 8000; % Sampling frequency 8kHz
fc = 800; % Passband cutoff frequency 800Hz
fs = 1600; % Stopband frequency 1.6kHz
Rp = 1; % Passband ripple in dB
Rs = 20; % Stopband attenuation in dB
Wp = 2*fc/Fs; % Normalized passband edge frequency
Ws = 2*fs/Fs; % Normalized stopband edge frequency
%
SOFTWARE DEVELOPMENTS AND EXPERIMENTS USING THE TMS320C55X 285
% Filter design
%
[N,Wn]= buttord(Wp,Ws,Rp,Rs); % Filter order selection
[b,a]= butter(N,Wn); % Butterworth filter design
[Z,P,K]= tf2zp(b,a); % Transfer function to zero-pole
[sos,G]= zp2sos(Z,P,K); % Zero-pole to second-order section
This program generates a fourth-order IIR filter with the following coefficient vectors:
b = [0.0098,0.0393,0.0590,0.0393,0.0098]
a = [1.0000,÷1.9908,1.7650,÷0.7403,0.1235].
The fourth-order IIR filter is then converted into two second-order sections repre-
sented by the coefficients matrix sos and the overall system gain G. By decomposing the
coefficient matrix sos defined in (6.4.19), we obtain two matrices
b =
0:0992 0:1984 0:0992
0:0992 0:1984 0:0992
¸
and a =
1:0 ÷0:8659 0:2139
1:0 ÷1:1249 0:5770
¸
, (6:7:1)
where we equally distribute the overall gain factor into each second-order cascade
configuration for simplicity. In the subsequent sections, we will use this Butterworth
filter for the TMS320C55x experiments.
6.7.2 Experiment 6A ± Floating-Point C Implementation
For an IIR filter consists of K second-order sections, the I/O equation of the cascade
direct-form II realization is given by Equation (6.4.14). The C implementation of
general cascade second-order sections can be written as follows:
temp = input[n];
for(k = 0; k < IIR_SECTION; k÷÷)
{
w[k][0]= temp ÷ a[k][1]*w[k][1]÷ a[k][2]*w[k][2];
temp = b[k][0]*w[k][0]+ b[k][1]*w[k][1]+ b[k][2]*w[k][2];
w[k][2]= w[k][1]; /* w(n÷2) <- w(n÷1) */
w[k][1]= w[k][0]; /* w(n÷1) <- w(n) */
}
output[n]= temp;
where a[][]and b[][]are filter coefficient matrices defined in (6.7.1), and w[][]is the
signal buffer for w
k
(n ÷m), m = 0, 1, 2. The row index k represents the kth second-
order IIR filter section, and the column index points at the filter coefficient or signal
sample in the buffer.
As mentioned in Chapter 5, the zero-overhead repeat loop, multiply±accumulate
instructions, and circular buffer addressing modes are three important features of
DSP processors. To better understand these features, we write the IIR filter function
in C using data pointers to simulate the circular buffers instead of two-dimensional
arrays. We also arrange the C statements to mimic the DSP multiply/accumulate
operations. The following C program is an IIR filter that consists of Ns second-order
286 DESIGN AND IMPLEMENTATION OF IIR FILTERS
sections in cascade form. The completed block IIR filter function (iir.c) written in
floating-point C language is provided in the experimental software package.
m = Ns *5; /* Setup for circular buffer C[m] */
k = Ns *2; /* Setup for circular buffer w[k] */
j = 0;
w_0 = x[n]; /* Get input signal */
for (i = 0; i < Ns; i÷÷)
{
w_0 ÷= *(w÷l)* *(C÷j); j÷÷; l =(l÷Ns)%k;
w_0 ÷= *(w÷l)* *(C÷j); j÷÷;
temp = *(w÷l);
*(w÷l) = w_0;
w_0 = temp * *(C÷j); j÷÷;
w_0 ÷= *(w÷l) * *(C÷j); j÷÷; l =(l÷Ns)%k;
w_0 ÷= *(w÷l) * *(C÷j); j =(j÷1)%m; l =(l÷1)%k;
}
y[n] = w_0 /* Save output */
The coefficient and signal buffers are configured as circular buffers shown in Figure
6.25. The signal buffer contains two elements, w
k
(n ÷1) and w
k
(n ÷2), for each second-
order section. The pointer address is initialized pointing at the first sample w
1
(n ÷1)
in the buffer. The coefficient vector is arranged with five coefficients (a
1k
, a
2k
, b
2k
, b
0k
,
Coefficient
buffer C[] w[]
Signal
buffer
offset =
number of
sections
:
a
11
a
21
b
21
b
01
b
11
a
12
a
22
b
22
b
02
:
b
12
Section 1
coefficients
Section 2
coefficients
a
1K
a
2K
b
2K
b
0K
b
1K
Section K
coefficients
w
1
(n−1)
:
:
w
K
(n−1)
w
2
(n−1)
w
1
(n−2)
:
:
w
2
(n−2)
w
K
(n−2)
Figure 6.25 IIR filter coefficient and signal buffers configuration
SOFTWARE DEVELOPMENTS AND EXPERIMENTS USING THE TMS320C55X 287
and b
1k
) per section with the coefficient pointer initialized pointing at the first coeffi-
cient, a
11
. The circular pointers are updated by j =(j÷1)%m and l =(l÷1)%k, where m
and k are the sizes of the coefficient and signal buffers, respectively.
The C function exp6a.c used for Experiment 6A is listed in Table 6.3. This program
calls the software signal generator signal_gen2()to create a block of signal samples
for testing. It then calls the IIR filter function iir to perform the lowpass filtering
process. The lowpass filter used for the experiment is the fourth-order Butterworth IIR
Table 6.3 List of floating-point C implementation of IIR filter
/*
exp6a.c ÷ Direct-form II IIR function implementation
in floating-point C and using signal generator
*/
#define M 128 /* Number of samples per block */
#define Ns 2 /* Number of second-order sections */
/* Low-pass IIR filter coefficients */
float C[Ns*5]= { /* i is section index */
/* A [i][1],A [i][2],B [i][2],B [i][0],B [i][1] */
÷0.8659, 0.2139, 0.0992, 0.0992, 0.1984,
÷1.1249, 0.5770, 0.0992, 0.0992, 0.1984};
/* IIR filter signal buffer:
w[]= w[i][n÷1],w[i÷1][n÷1],...,w[i][n÷2],w[i÷1][n÷2],... */
float w[Ns*2];
int out[M];
int in[M];
/* IIR filter function */
extern void iir(int *, int, int *, float *, int, float *);
/* Software signal generator */
extern void signal_gen2(int *, int);
void main(void)
{
int i;
/* Initialize IIR filter signal buffer */
for(i = 0; i < Ns*2;i÷÷)
w[i]= 0;
/* IIR filtering */
for(;;)
{
signal_gen2(in, M); /* Generate a block of samples */
iir(in, M, out, C, Ns, w); /* Filter a block of samples */
}
}
288 DESIGN AND IMPLEMENTATION OF IIR FILTERS
filter designed in Section 6.7.1. We rearranged the filter coefficients for using circular
buffer. Two temporary variables, temp and w_0, are used for intermediate storage.
Go through the following steps for Experiment 6A:
1. Create the project exp6a and add the linker command file exp6.cmd, the C
functions iir.c, epx6a.c, signal_gen2.c and sine.asm to the project.
The lowpass filter iir()will attenuate high-frequency components from the input
signal generated by the signal generator signal_gen2(), which uses the recursive
sinewave generator sine()to generate three sinewaves at 800 Hz, 1.8 kHz, and
3.3 kHz.
2. Use rts55.lib for initializing the C function main() and build the project
exp6a.
3. Set a breakpoint at the statement for(;;)of the main()function, and use the CCS
graphic function to view the 16-bit integer output samples in the buffer out[]. Set
data length to 128 for viewing one block of data at a time. Animate the filtering
process, and observe the filter output as a clean 800 Hz sinewave.
4. Profile the IIR filter performance by measuring the average DSP clock cycles.
Record the clock cycles and memory usage of the floating-point C implementation.
5. Overflow occurs when the results of arithmetic operations are larger than the fixed-
point DSP can represent. Before we move on to fixed-point C implementation, let us
examine the IIR filter for possible overflow. First, change the conditional compiling
bit CHECK_OVERFLOW defined in iir.c from 0 to 1 to enable the sections that
search for maximum and minimum intermediate values. Then, add w_max and
w_min to the CCS watch window. Finally, run the experiment in the animation
mode and examine the values of w_max and w_min. If [w_max[ _ 1 or [w_min[ _ 1,
an overflow will happen when this IIR filter is implemented by a 16-bit fixed-point
processor. If the overflow is detected, modify the IIR filter routine by scaling down
the input signal until the values [w_max[ and [w_min[ are less than 1. Remember to
scale up the filter output if the input is scaled down.
6.7.3 Experiment 6B ± Fixed-Point C Implementation Using Intrinsics
Since the TMS320C55x is a fixed-point device, the floating-point implementation of the
IIR filter given in Experiment 6A is very inefficient for real-time applications. This is
because the floating-point implementation needs the floating-point math library func-
tions that use fixed-point hardware to realize floating-point operations. In order to
improve the performance, fixed-point C implementation should be considered. To ease
the burden of fixed-point C programming, the TMS320C55x C compiler provides a
set of intrinsics to handle specific signal processing operations.
The C55x intrinsics produce assembly language statements directly into the pro-
grams. The intrinsics are specified with a leading underscore and are used as C func-
SOFTWARE DEVELOPMENTS AND EXPERIMENTS USING THE TMS320C55X 289
tions. The intrinsic function names are similar to their mnemonic assembly counter-
parts. For example, the following signed multiply±accumulate intrinsic
z = _smac(z,x,y); /* Perform signed z = z÷x*y */
will perform the equivalent assembly instruction
macm Xmem,Ymem,ACx
Table 6.4 lists the intrinsics supported by the TMS320C55x.
Table 6.4 Intrinsics provided by the TMS320C55x C compiler
C Compiler Intrinsic
(a,b,c are 16-bit and d,e,f are 32-bit data) Description
c = _sadd(int a, int b); Adds 16-bit integers a and b, with SATA set,
producing a saturated 16-bit result c.
f = _lsadd(long d, long e); Adds 32-bit integers d and e, with SATD set,
producing a saturated 32-bit result f.
c = _ssub(int a, int b); Subtracts 16-bit integer b from a with SATA
set, producing a saturated 16-bit result c.
f = _lssub(long d, long e); Subtracts 32-bit integer e from d with SATD
set, producing a saturated 32-bit result f.
c = _smpy(int a, int b); Multiplies a and b, and shifts the result left
by 1. Produces a saturated 16-bit result c.
(upper 16-bit, SATD and FRCT set)
f = _lsmpy(int a, int b); Multiplies a and b, and shifts the result left
by 1. Produces a saturated 32-bit result f.
(SATD and FRCT set)
f = _smac(long d, int a, int b); Multiplies a and b, shifts the result left by 1,
and adds it to d. Produces a saturated
32-bit result f. (SATD, SMUL and FRCT
set)
f = _smas(long d, int a, int b); Multiplies a and b, shifts the result left by 1,
and subtracts it from d. Produces a 32-bit
result f. (SATD, SMUL and FRCT set)
c = _abss(int a); Creates a saturated 16-bit absolute value.
c = [a[, _abss(0x8000) => 0x7FFF (SATA
set)
f = _labss(long d); Creates a saturated 32-bit absolute value.
f = [d[, _labss(0x8000000) => 0x7FFFFFFF
(SATD set)
c = _sneg(int a); Negates the 16-bit value with saturation.
c = ÷ a, _sneg(0xffff8000) => 0x00007FFF
290 DESIGN AND IMPLEMENTATION OF IIR FILTERS
Table 6.4 (continued )
C Compiler Intrinsic
(a,b,c are 16-bit and d,e,f are 32-bit data) Description
f = _lsneg(long d); Negates the 32-bit value with saturation.
f =÷d, _lsneg(0x80000000) =>
0x7FFFFFFF
c = _smpyr(int a, int b); Multiplies a and b, shifts the result left by 1,
and rounds the result c. (SATDand FRCTset)
c = _smacr(long d, int a, int b); Multiplies a and b, shifts the result left by 1,
adds the result to d, and then rounds the
result c. (SATD, SMUL and FRCT set)
c = _smasr(long d, int a, int b); Multiplies a and b, shifts the result left by 1,
subtracts the result from d, and then rounds
the result c. (SATD, SMUL and FRCT set)
c = _norm(int a); Produces the number of left shifts needed to
normalize a and places the result in c.
c = _lnorm(long d); Produces the number of left shifts needed to
normalize d and places the result in c.
c = _rnd(long d); Rounds d to produces the 16-bit saturated
result c. (SATD set)
c = _sshl(int a, int b); Shifts a left by b and produces a 16-bit result
c. The result is saturated if b is greater than or
equal to 8. (SATD set)
f = _lsshl(long d, int a); Shifts a left by b and produces a 32-bit result
f. The result is saturated if a is greater than or
equal to 8. (SATD set)
c = _shrs(int a, int b); Shifts a right by b and produces a 16-bit result
c. (SATD set)
f = _lshrs(long d, int a); Shifts d right by a and produces a 32-bit result
f. (SATD set)
c = _addc(int a, int b); Adds a, b, and Carry bit and produces a
16-bit result c.
f = _laddc(long d, int a); Adds d, a, and Carry bit and produces a
32-bit result f.
The floating-point IIR filter function given in the previous experiment can be con-
verted to the fixed-point C implementation using these intrinsics. To prevent inter-
mediate overflow, we scale the input samples to Q14 format in the fixed-point
implementation. Since the largest filter coefficient is between 1 and 2, we use Q14
representation for the fixed-point filter coefficients defined in (6.7.1). The implementa-
tion of the IIR filter in the fixed-point Q14 format is given as follows:
SOFTWARE DEVELOPMENTS AND EXPERIMENTS USING THE TMS320C55X 291
m = Ns*5; /* Setup for circular buffer coef[m] */
k = Ns*2; /* Setup for circular buffer w[k] */
j = 0;
w_0 = (long)x[n]<14; /* Q14 input (scaled)*/
for(i = 0; i < Ns; i÷÷)
{
w_0 = _smas(w_0,*(w÷l),*(coef÷j)); j÷÷; l =(l÷Ns)%k;
w_0 = _smas(w_0,*(w÷l),*(coef÷j)); j÷÷;
temp = *(w÷l);
*(w÷l) = w_015;
w_0 = _lsmpy(temp,*(coef÷j)); j÷÷;
w_0 = _smac(w_0,*(w÷l), *(coef÷j)); j÷÷; l =(l÷Ns)%k;
w_0 = _smac(w_0,*(w÷l), *(coef÷j)); j =(j÷1)%m; l =(l÷1)%k;
}
y[n]= w_014; /* Q15 output */
Go through the following steps for Experiment 6B:
1. Create the project exp6b that include the linker command file exp6.cmd, the C
functions exp6b.c, iir_i1.c, signal_gen2.c, and the assembly routine
sine.asm.
2. Use rts55.lib for initializing the C function main(), and build the project
exp6b.
3. Set a breakpoint at the statement for(;;)of the main()function, and use the CCS
graphic function to view the 16-bit integer output samples in the buffer out[]. Set
data length to 128 for viewing one block of data at a time. Animate the filtering
process and observe the filter output as a clean 800 Hz sinewave.
4. Profile the IIR filter function iir_i1(), and compare the results with those
obtained in Experiment 6A.
5. In file iir_i1(), set the scaling factor SCALE to 0 so the samples will not be scaled.
Rebuild the project, and run the IIR filter experiment in the animation mode. We
will see the output distortions caused by the intermediate overflow.
6.7.4 Experiment 6C ± Fixed-Point C Programming Considerations
From the previous experiments, the fixed-point IIR filter implementation using intrin-
sics has greatly improved the efficiency of the fixed-point IIR filter using the floating-
point implementation. We can further enhance the C function performance by taking
advantage of the compiler optimization and restructuring the program to let the C
compiler generate a more efficient assembly code.
292 DESIGN AND IMPLEMENTATION OF IIR FILTERS
Run-time support library functions
The TMS320C55x C compiler has many built-in C functions in its run-time support
library rts55.lib. Although these functions are helpful, most of them may run at a
slower speed due to the nested library function calls. We should try to avoid using them
in real-time applications if possible. For example, the MOD(%)operation we use to
simulate the circular addressing mode can be replaced by a simple AND(&)operation if
the size of the buffer is a base 2 number, such as 2, 4, 8, 16, and so on. The example given
in Table 6.5 shows the compiler will generate a more efficient assembly code (by
avoiding calling the library function I$$MOD) when using the logic operator AND
than the MOD operator, because the logic operation AND does not invoke any
function calls.
Loop counters
The for-loop is the most commonly used loop control operations in C programs for
DSP applications. The assembly code generated by the C55x C compiler varies depend-
ing on how the for-loop is written. Because the compiler must verify both the positive
and negative conditions of the integer loop counter against the loop limit, it creates
more lines of assembly code to check the entrance and termination of the loop. By using
an unsigned integer as a counter, the C compiler only needs to generate a code that
compares the positive loop condition. Another important loop-control method is to use
a down counter instead of an up counter if possible. This is because most of the built-in
conditional instructions act upon zero conditions. The example given in Table 6.6 shows
the assembly code improvement when it uses an unsigned integer as a down counter.
Local repeat loop
Using local repeat-loop is another way to improve the DSP run-time efficiency. The
local repeat-loop uses the C55x instruction-buffer-queue (see Figure 2.2) to store all the
Table 6.5 Example to avoid using library function when applying modulus operation
;Ns = 2; ;Ns = 2;
;k = 2*Ns; ;k = 2*Ns÷1;
;l =(l÷Ns)%k; ;l =(l÷Ns)&k;
MOV *SP(#04h),AR1 MOV T0,*SP(#07h)
MOV *SP(#0ah),T1 ADD *SP(#0bh),AR1
ADD *SP(#0bh),AR1,T0 AND *SP(#0ah),AR1
CALL I$$MOD MOV AR1,*SP(#0bh)
MOV T0,*SP(#0bh)
SOFTWARE DEVELOPMENTS AND EXPERIMENTS USING THE TMS320C55X 293
Table 6.6 Example of using unsigned integer as down counter for loop control
;int i ;unsigned int i
;for(i = 0; i<Ns; i÷÷) ;for(i = Ns; i>0; i÷÷)
;{ ;{
... ... ... ...
;} ;}
MOV *SP(#04h),AR1 MOV *SP(#04h),AR1
MOV #0,*SP(#06h) MOV AR1,*SP(#06h)
MOV *SP(#06h),AR2 BCC L3,AR1 == #0
CMP AR2 >= AR1, TC1
BCC L3,TC1
... ...
... ... ADD #-1,*SP(#06h)
MOV *SP(#06h),AR1
MOV *SP(#04h),AR2
BCC L2,AR1 != #0
ADD #1,*SP(#06h)
MOV *SP(#06h),AR1
CMP AR1 < AR2, TC1
BCC L2,TC1
instructions within a loop. Local repeat-loop can execute the instructions repeatedly
within the loop without additional instruction fetches. To allow the compiler to generate
local-repeat loops, we should reduce the number of instructions within the loop because
the size of instruction buffer queue is only 64 bytes.
Compiler optimization
The C55x C compiler has many options. The -on option (n = 0, 1, 2, or 3) controls the
compiler optimization level of the assembly code it generated. For example, the -o3
option will perform loop optimization, loop unrolling, local copy/constant propagation,
simplify expression statement, allocate variables to registers, etc. The example given in
Table 6.7 shows the code generated with and without the -o3 optimization.
Go through the following steps for Experiment 6C:
1. Create the project exp6c, add the linker command file exp6.cmd, the C functions
exp6c.c, iir_i2.c, signal_gen2.c, and the assembly routine sine.asm
into the project. The C function iir_i2.c uses unsigned integers for loop coun-
ters, and replaces the MOD operation with AND operation for the signal buffer.
2. Use rts55.lib for initializing the C function main(), and build the project
exp6c.
3. Relocate the C program and data variables into SARAM and DARAM sections
defined by the linker command files. Use pragma to allocate the program and data
memory as follows:
294 DESIGN AND IMPLEMENTATION OF IIR FILTERS
Table 6.7 Example of compiler without and with -o3 optimization option
-o3 disabled -o3 enabled
;Ns = 2; ;Ns = 2;
;for(i = Ns*2; i > 0; i÷÷) ;for(i = Ns*2; i > 0; i÷÷)
; *ptr÷÷ = 0; ; *ptr÷÷ = 0;
MOV #4,*SP(#00) RPT #3
MOV *SP(#00),AR1 MOV #0,*AR3÷
BCC L2, AR1 == #0
MOV *SP(#01h),AR3
ADD #1,AR3,AR1
MOV AR1,*SP(#01h)
MOV #0,*AR3
ADD #-1,*SP(#00h)
MOV *SP(#00h),AR1
BCC L1,AR1 != #0
± Place the main()and iir()functions into the program SARAM, and name the
section iir_code.
± Allocate the input and output buffers in []and out []to data SARAM, and
name the sections input and output.
± Put the IIR filter coefficient buffer C []in a separate data SARAM section, and
name the section iir_coef.
± Place the temporary buffer w []and temporary variables in a DARAM section,
and name it iir_data.
4. Enable -o3 and -mp options to rebuild the project.
5. Set a breakpoint at the statement for(;;)of the main()function, and use the CCS
graphic function to view the 16-bit integer output samples in the buffer out []. Set
data length to 128 for viewing one block of data at a time. Animate the filtering
process and observe the filter output as a clean 800 Hz sinewave.
6. Profile the IIR filter iir_i2(), and compare the result with those obtained in
Experiment 6B.
6.7.5 Experiment 6D ± Assembly Language Implementations
The advantages of C programming are quick prototyping, portable to other DSP
processors, easy maintenance, and flexibility for algorithm evaluation and analysis.
However, the IIR filter implementations given in previous experiments can be more
efficient if the program is written using the C55x assembly instructions. By examining
SOFTWARE DEVELOPMENTS AND EXPERIMENTS USING THE TMS320C55X 295
the IIR filter function used for Experiments 6B and 6C, we anticipate that the filter
inner-loop can be implemented by the assembly language in seven DSP clock cycles.
Obviously, from the previous experiments, the IIR filter implemented in C requires
more cycles. To get the best performance, we can write the IIR filtering routine in
assembly language. The trade-off between C and assembly programming is the time
needed as well as the difficulties encountered for program development, maintenance,
and system migration from one DSP system to the others.
The IIR filter realized by cascading the second-order sections can be implemented in
assembly language as follows:
masm *AR3÷,*AR7÷,AC0 ; AC0 ÷= AC0÷a1*w(n÷1)
masm T3 = *AR3,*AR7÷,AC0 ; AC0 ÷= AC0÷a2*w(n÷2)
mov rnd(hi(AC0)),*AR3÷ ; Update w(n)buffer
mpym *AR7÷,T3,AC0 ; AC0 = bi2*w(n-2)
macm *(AR3÷T1), *AR7÷,AC0 ; AC0 ÷= AC0÷bi0*w(n)
macm *AR3÷,*AR7÷,AC0 ; AC0 ÷= AC0÷bi1*w(n÷1)
mov rnd(hi(AC0)),*AR1÷ ; Store result
The assembly program uses three pointers. The IIR filter signal buffer for w
k
(n) is
addressed by the auxiliary register AR3, while the filter coefficient buffer is pointed at
by AR7. The filter output is rounded and placed in the output buffer pointed at by AR1.
The code segment can be easily modified for filtering either a single sample of data or a
block of samples. The second-order IIRfilter sections can be implemented using the inner
repeat loop, while the outer loop can be used for controlling samples in blocks. The input
sample is scaled down to Q14 format, and the IIRfilter coefficients are also represented in
Q14 format to prevent overflow. To compensate the Q14 format of coefficients and signal
samples, the final result y(n) is multiplied by 4 (implemented by shifting two bits to the
left) to scale it back to Q15 format and store it with rounding. Temporary register T3 is
used to hold the second element w
k
(n ÷2) when updating the signal buffer.
Go through the following steps for Experiment 6D:
1. Create the project exp6d, and include the linker command file exp6.cmd, the
C functions exp6d.c, signal_gen2.c, the assembly routine sine.asm, and
iirform2.asm into the project. The prototype of the IIR filter routine is written as
void iirform2(int *x, unsigned int M, int *y, int *h,
unsigned int N, int *w);
where
x is the pointer to the input data buffer in []
h is the pointer to the filter coefficient buffer C []
y is the pointer to the filter output buffer out []
w is the pointer to the signal buffer w []
M is the number of samples in the input buffer
N is the number of second-order IIR filter sections.
2. Use rts55.lib for initializing the C function main(), and build the project
exp6d.
296 DESIGN AND IMPLEMENTATION OF IIR FILTERS
3. Set a breakpoint at the statement for(;;)of the main()function, and use the CCS
graphic function to view the 16-bit integer output samples in the buffer out[]. Set
data length to 128 for viewing one block of data at a time. Animate the filtering
process, and observe the filter output as a clean 800 Hz sinewave.
4. Profile the IIR filter iirform2(), and compare the profile result with those
obtained in Experiments 6B and 6C.
References
[1] N. Ahmed and T. Natarajan, Discrete-Time Signals and Systems, Englewood Cliffs, NJ: Prentice-
Hall, 1983.
[2] V. K. Ingle and J. G. Proakis, Digital Signal Processing Using MATLAB V.4, Boston: PWS
Publishing, 1997.
[3] Signal Processing Toolbox for Use with MATLAB, The Math Works Inc., 1994.
[4] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, Englewood Cliffs, NJ:
Prentice-Hall, 1989.
[5] S. J. Orfanidis, Introduction to Signal Processing, Englewood Cliffs, NJ: Prentice-Hall, 1996.
[6] J. G. Proakis and D. G. Manolakis, Digital Signal Processing ± Principles, Algorithms, and
Applications, 3rd Ed., Englewood Cliffs, NJ: Prentice-Hall, 1996.
[7] S. K. Mitra, Digital Signal Processing: A Computer-Based Approach, New York, NY: McGraw-
Hill, 1998.
[8] D. Grover and J. R. Deller, Digital Signal Processing and the Microcontroller, Englewood Cliffs,
NJ: Prentice-Hall, 1999.
[9] F. Taylor and J. Mellott, Hands-On Digital Signal Processing, New York, NY: McGraw-Hill,
1998.
[10] S. D. Stearns and D. R. Hush, Digital Signal Analysis, 2nd Ed., Englewood Cliffs, NJ: Prentice-
Hall, 1990.
[11] S. S. Soliman and M. D. Srinath, Continuous and Discrete Signals and Systems, 2nd Ed., Engle-
wood Cliffs, NJ: Prentice-Hall, 1998.
[12] L. B. Jackson, Digital Filters and Signal Processing, 2nd Ed., Boston, MA: Kluwer Academic
Publishers, 1989.
Exercises
Part A
1. Find the Laplace transform of
x(t) = e
÷at
sin(
0
t)u(t):
2. Find the inverse Laplace transform of
(a) X(s) =
2s ÷1
s
3
÷3s
2
÷4s
:
(b) X(s) =
2s
2
÷3s
s
3
÷4s
2
÷5s ÷2
:
(c) X(s) =
s ÷3
s
2
÷4s ÷13
:
EXERCISES 297
3. Given the transfer function of a continuous-time system as
H(s) =
s(s ÷5)(s
2
÷s ÷1)
(s ÷1)(s ÷2)(s ÷3)(s
2
÷cs ÷5)
, c _ 0:
(a) Show a plot of the poles and zeros of H(s).
(b) Discuss the stability of this system for the cases c = 0 and c > 0.
4. Consider the circuit shown in Figure 6.3 with R = 1 and C = 1F. The input signal to the
circuit is expressed as
x(t) =
1 0 _ t _ 1
0 elsewhere.

Show that the output is
y(t) = (1 ÷e
÷t
)u(t) ÷[1 ÷e
÷(t÷1)
[u(t ÷1):
5. Given the transfer function
H(s) =
1
(s ÷1)(s ÷2)
,
find the H(z) using the impulse-invariant method.
6. Given the transfer function of an analog IIR notch filter as
H(s) =
s
2
÷1
s
2
÷s ÷1
,
design a digital filter using bilinear transform with notch frequency 100 Hz and sampling rate
1 kHz.
7. Given an analog IIR bandpass filter that has resonance at 1 radian/second with the transfer
function
H(s) =
5s ÷1
s
2
÷0:4s ÷1
,
design a digital resonant filter that resonates at 100 Hz with the sampling rate at 1 kHz.
8. Design a second-order digital Butterworth filter using bilinear transform. The cut-off
frequency is 1 kHz at a sampling frequency of 10 kHz.
9. Repeat the previous problem for designing a highpass filter with the same specifications.
10. Design a second-order digital Butterworth bandpass filter with the lower cut-off frequency
200 Hz, upper cut-off frequency 400 Hz, and sampling frequency 2000 Hz.
11. Design a second-order digital Butterworth bandstop filter that has the lower cut-off fre-
quency 200 Hz, upper cut-off frequency 400 Hz, and sampling frequency 2000 Hz.
12. Given the transfer function
H(z) =
0:5(z
2
÷1:1z ÷0:3)
z
3
÷2:4z
2
÷1:91z ÷0:504
,
find the following realizations:
298 DESIGN AND IMPLEMENTATION OF IIR FILTERS
(a) Direct-form II.
(b) Cascade of first-order sections.
(c) Parallel form in terms of first-order sections.
13. Given an IIR filter transfer function
H(z) =
(3 ÷2:1z
÷1
)(2:7 ÷4:2z
÷1
÷5z
÷2
)
(1 ÷0:52z
÷1
)(1 ÷z
÷1
÷0:34z
÷2
)
,
realize the transfer function in
(a) cascade canonical form, and
(b) parallel form.
14. Draw the direct-form I and II realizations of the transfer function
H(z) =
(z
2
÷2z ÷2)(z ÷0:6)
(z ÷0:8)(z ÷0:8)(z
2
÷0:1z ÷0:8)
:
15. Given a third-order IIR filter transfer function
H(z) =
0:44z
2
÷0:36z ÷0:02
z
3
÷0:4z
2
÷0:18z ÷0:2
,
find and draw the following realizations:
(a) direct-form II,
(b) cascade realization based on direct-form II realization of each section, and
(c) parallel direct-form II realization.
16. The difference filter is defined as
y(n) = b
0
x(n) ÷b
1
x(n ÷1),
derive the frequency response of the filter and show that this is a crude highpass filter.
17. Given an IIR filter with transfer function
H(z) =
(1 ÷1:414z
÷1
÷z
÷2
)(1 ÷2z
÷1
÷z
÷2
)
(1 ÷0:8z
÷1
÷0:64z
÷2
)(1 ÷1:0833z
÷1
÷0:25z
÷2
)
,
(a) find the poles and zeros of the filter,
(b) using the stability triangle to check if H(z) is a stable filter.
18. Consider the second-order IIR filter with the I/O equation
y(n) = x(n) ÷a
1
y(n ÷1) ÷a
2
y(n ÷2); n _ 0,
where a
1
and a
2
are constants.
(a) Find the transfer function H(z).
(b) Discuss the stability conditions related to the cases:
(1)
a
2
1
4
÷a
2
< 0.
EXERCISES 299
(2)
a
2
1
4
÷a
2
> 0.
(3)
a
2
1
4
÷a
2
= 0.
19. An allpass filter has a magnitude response that is unity for all frequencies, that is,
[H(!)[ = 1 for all !: Such filters are useful for phase equalization of IIR designs. Show
that the transfer function of an allpass filter is of the form
H(z) =
z
÷L
÷b
1
z
÷L÷1
÷ ÷b
L
1 ÷b
1
z
÷1
÷ ÷b
L
z
÷L
,
where all coefficients are real.
20. A first-order allpass filter has the transfer function
H(z) =
z
÷1
÷a
1 ÷az
÷1
:
(a) Draw the direct-form I and II realizations.
(b) Show that [H(!)[ = 1 for all !.
(c) Sketch the phase response of this filter.
21. Design a second-order resonator with peak at 500 Hz, bandwidth 32 Hz, and operating at the
sampling rate 10 kHz.
Part B
22. Given the sixth-order IIR transfer function
H(z) =
6 ÷17z
÷1
÷33z
÷2
÷25z
÷3
÷20z
÷4
÷5z
÷5
÷8z
÷6
1 ÷2z
÷1
÷3z
÷2
÷z
÷3
÷0:2z
÷4
÷0:3z
÷5
÷0:2z
÷6
,
find the factored form of the IIR transfer function in terms of second-order section using
MATLAB.
23. Given the fourth-order IIR transfer function
H(z) =
12 ÷2z
÷1
÷3z
÷2
÷20z
÷4
6 ÷12z
÷1
÷11z
÷2
÷5z
÷3
÷z
÷4
,
(a) using MATLAB to express H(z) in factored form,
(b) develop two different cascade realizations, and
(c) develop two different parallel realizations.
24. Design and plot the magnitude response of an elliptic IIR lowpass filter with the following
specifications using MATLAB:
Passband edge at 800 Hz
Stopband edge at 1000 Hz
Passband ripple of 0.5 dB
300 DESIGN AND IMPLEMENTATION OF IIR FILTERS
Minimum stopband attenuation of 40 dB
Sampling rate of 4 kHz.
25. Design an IIR Butterworth bandpass filter with the following specifications:
Passband edges at 450 Hz and 650 Hz
Stopband edges at 300 Hz and 750 Hz
Passband ripple of 1 dB
Minimum stopband attenuation of 40 dB
Sampling rate of 4 kHz.
26. Design a type I Chebyshev IIR highpass filter with passband edge at 700 Hz, stopband edge
at 500 Hz, passband ripple of 1 dB, and minimum stopband attenuation of 32 dB. The
sampling frequency is 2 kHz. Plot the magnitude response of the design filter.
27. Given an IIR lowpass filter with transfer function
H(z) =
0:0662(1 ÷3z
÷1
÷3z
÷2
÷z
÷3
)
1 ÷0:9356z
÷1
÷0:5671z
÷2
÷0:1016z
÷3
,
(a) plot the first 32 samples of the impulse response using MATLAB,
(b) filter the input signal that consists of two sinusoids of normalized frequencies 0.1 and 0.8
using MATLAB.
28. It is interesting to examine the frequency response of the second-order resonator filter given
in (6.6.15) as the radius r
p
and the pole angle !
0
are varied. Using the MATLAB to compute
and plot
(a) The magnitude response for !
0
= p=2 and various values of r
p
.
(b) The magnitude response for r
p
= 0:95 and various values of !
0
.
Part C
29. An IIR filter design and implementation using the direct-form II realization.
(a) Use MATLAB to design an elliptic bandpass IIR filter that meets the following speci-
fications:
± Sampling frequency is 8000 Hz
± Lower stopband extends from 0 to 1200 Hz
± Upper stopband extends from 2400 to 4000 Hz
± Passband starts from 1400 Hz with bandwidth of 800 Hz
± Passband ripple should be no more than 0.3 dB
± Stopband attenuation should be at least 30 dB.
(b) For the elliptic bandpass IIR filter obtained above,
± plot the amplitude and phase responses
EXERCISES 301
± realize the filter using the cascade of direct-form II second-order sections.
(c) Implement the filter using floating-point C language and verify the filter implementation.
(d) Modify the C program using the C55x intrincis.
(e) Implement the filter in C55x assembly language.
(f) Using the CCS to show the filter results in both time domain and frequency domain.
(g) Profile the different implementations (C, intrinsics, and assembly) of the elliptic IIR filter
and compare the results.
30. The overflow we saw in Experiments 6A and 6B is called the intermediate overflow. It
happens when the signal buffer of the direct-form II realization uses 16-bit wordlength.
Realizing the IIR filter using the direct-form I structure can eliminate the intermediate
overflow by keeping the intermediate results in the 40-bit accumulators. Write an assembly
routine to realize the fourth-order lowpass Butterworth IIR filter in the direct-form I
structure.
31. Implement the recursive quadrature oscillator shown in Figure 6.23 in TMS320C55x assem-
bly language.
32. Verify the IIR filter design in real time using a C55x EVM/DSK. Use a signal generator and
a spectrum analyzer to measure the amplitude response and plot it. Evaluate the IIR filter
according to the following steps:
± Set the TMS320C55x EVM/DSK to 8 kHz sampling rate
± Connect the signal generator output to the audio input of the EVM/DSK
± Write an interrupt service routine (ISR) to handle input samples
± Process the random samples at 128 samples per input block
± Verify the filter using a spectrum analyzer.
302 DESIGN AND IMPLEMENTATION OF IIR FILTERS
7
Fast Fourier Transform and Its
Applications
Frequency analysis of digital signals and systems was discussed in Chapter 4. To per-
form frequency analysis on a discrete-time signal, we converted the time-domain
sequence into the frequency-domain representation using the z-transform, the
discrete-time Fourier transform (DTFT), or the discrete Fourier transform (DFT). The
widespread application of the DFT to spectral analysis, fast convolution, and data
transmission is due to the development of the fast Fourier transform (FFT) algorithm
for its computation. The FFT algorithm allows a much more rapid computation of the
DFT, was developed in the mid-1960s by Cooley and Tukey.
It is critical to understand the advantages and the limitations of the DFT and how to
use it properly. We will discuss the important properties of the DFT in Section 7.1. The
development of FFT algorithms will be covered in Section 7.2. In Section 7.3, we will
introduce the applications of FFTs. Implementation considerations such as computa-
tional issues and finite-wordlength effects will be discussed in Section 7.4. Finally,
implementation of the FFTalgorithmusing the TMS320C55x for experimental purposes
will be given in Section 7.5.
7.1 Discrete Fourier Transform
As discussed in Chapter 4, we perform frequency analysis of a discrete-time signal x(n)
using the DTFT defined in (4.4.4). However, X(!) is a continuous function of frequency
! and the computation requires an infinite-length sequence x(n). Thus the DTFT cannot
be implemented on digital hardware. We define the DFT in Section 4.4.3 for N samples
of x(n) at N discrete frequencies. The input to the N-point DFT is a digital signal
containing N samples and the output is a discrete-frequency sequence containing N
samples. Therefore the DFT is a numerically computable transform and is suitable for
DSP implementations.
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
7.1.1 Definitions
Given the DTFT X(!), we take N samples over the full Nyquist interval, 0 _ ! < 2p, at
discrete frequencies !
k
= 2pk=N, k = 0, 1, . . . , N ÷1. This is equivalent to evaluating
X(!) at N equally spaced frequencies !
k
, with a spacing of 2p=N radians (or f
s
=N Hz)
between successive samples. That is,
X(!
k
) =
¸
·
n=÷·
x(n)e
÷j(2p=N)kn
=
¸
N÷1
n=0
¸
·
l=÷·
x(n ÷lN)
¸ ¸
e
÷j(2p=N)kn
=
¸
N÷1
n=0
x
p
(n)e
÷j(2p=N)kn
, k = 0, 1, . . . , N ÷1, (7:1:1a)
where
x
p
(n) =
¸
·
l=÷·
x(n ÷lN) (7:1:1b)
is a periodic signal with period N.
In general, the equally spaced frequency samples do not represent the original
spectrum X(!) uniquely when x(n) has infinite duration. Instead, these frequency
samples correspond to a periodic sequence x
p
(n) as shown in (7.1.1). When the sequence
x(n) has a finite length N, x
p
(n) is simply a periodic repetition of x(n). Therefore
the frequency samples X(!
k
), k = 0, 1, . . . , N ÷1 uniquely represent the
finite-duration sequence x(n). Since x(n) = x
p
(n) over a single period, a finite-duration
sequence x(n) of length N has the DFT defined as
X(k) = X(!
k
) = X
2pk
N

=
¸
N÷1
n=0
x(n)e
÷j(2p=N)kn
, k = 0, 1, . . . , N ÷1, (7:1:2)
where X(k) is the kth DFT coefficient and the upper and lower indices in the summation
reflect the fact that x(n) = 0 outside the range 0 _ n _ N ÷1. Strictly speaking, the DFT
is a mapping between an N-point sequence in the time domain and an N-point sequence in
the frequency domain that is applicable in the computation of the DTFT of periodic and
finite-length sequences.
Example 7.1: If the signals ¦x(n)¦ are real valued and N is an even number, we can
show that X(0) and X(N=2) are real values and can be computed as
X(0) =
¸
N÷1
n=0
x(n)
304 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
and
X(N=2) =
¸
N÷1
n=0
e
÷jpn
x(n) =
¸
N÷1
n=0
(÷1)
n
x(n):
The DFT defined in (7.1.2) can also be written as
X(k) =
¸
N÷1
n=0
x(n)W
kn
N
, k = 0, 1, . . . N ÷1, (7:1:3)
where
W
kn
N
= e
÷j
2p
N

kn
= cos
2pkn
N

÷j sin
2pkn
N

, 0 _ k, n _ N ÷1 (7:1:4)
are the complex basis functions, or twiddle factors of the DFT. Each X(k) can be viewed
as a linear combination of the sample set ¦x(n)¦ with the coefficient set ¦W
kn
N
¦. Thus we
have to store the twiddle factors in terms of real and imaginary parts in the DSP
memory. Note that W
N
is the Nth root of unity since W
N
N
= e
÷j2p
= 1 = W
0
N
. All the
successive powers W
k
N
, k = 0, 1, . . . , N ÷1 are also Nth roots of unity, but in clockwise
direction on the unit circle. It can be shown that W
N=2
N
= e
÷jp
= ÷1, the symmetry
property
W
k÷N=2
N
= ÷W
k
N
, 0 _ k _ N=2 ÷1 (7:1:5a)
and the periodicity property
W
k÷N
N
= W
k
N
: (7:1:5b)
Figure 7.1 illustrates the cyclic property of the twiddle factors for an eight-point DFT.
W
8
0
1 =
3
8
7
8
W W − =
2
8
6
8
W W − =
1
8
5
8
W W − =
1
0
8
4
8
− = = −W W
3
8
W
2
8
W
1
8
W
Figure 7.1 Twiddle factors for DFT, N = 8 case
DISCRETE FOURIER TRANSFORM 305
The inverse discrete Fourier transform (IDFT) is used to transform the X(k) back into
the original sequence x(n). Given the frequency samples X(k), the IDFT is defined as
x(n) =
1
N
¸
N÷1
k=0
X(k)e
j(2p=N)kn
=
1
N
¸
N÷1
k=0
X(k)W
÷kn
N
, n = 0, 1, . . . , N ÷1: (7:1:6)
This is identical to the DFT with the exception of the normalizing factor 1/N and
the sign of the exponent of the twiddle factors. The IDFT shows that there is no loss
of information by transforming the spectrum X(k) back into the original time sequence
x(n). The DFT given in (7.1.3) is called the analysis transform since it analyzes the signal
x(n) at N frequency components. The IDFT defined in (7.1.6) is called the synthesis
transform because it reconstructs the signal x(n) from the frequency components.
Example 7.2: Consider the finite-length sequence
x(n) = a
n
, n = 0, 1, . . . , N ÷1,
where 0 < a < 1. The DFT of the signal x(n) is computed as
x(k) =
¸
N÷1
n=0
a
n
e
÷j(2pk=N)n
=
¸
N÷1
n=0
ae
÷j2pk=N

n
=
1 ÷(ae
÷j2pk=N
)
N
1 ÷ae
÷j2pk=N
=
1 ÷a
N
1 ÷ae
÷j2pk=N
, k = 0, 1, . . . , N ÷1:
The DFT and IDFT defined in (7.1.3) and (7.1.6), can be expressed in matrix±vector
form as
X = Wx (7:1:7a)
and
x =
1
N
W
+
X , (7:1:7b)
where x = [x(0) x(1) . . . x(N ÷1)[
T
is the signal vector, the frequency-domain DFT
coefficients are contained in the complex vector X = [X(0) X(1) . . . X(N ÷1)[
T
, and
the NxN twiddle-factor matrix (or DFT matrix) W is given by
W= W
kn
N

0_k, n_N÷1
=
1 1 1
1 W
1
N
W
N÷1
N
.
.
.
.
.
.
.
.
.
.
.
.
1 W
N÷1
N
W
(N÷1)
2
N

¸
¸
¸
¸
¸
¸
¸
¸
¸
,
(7:1:8)
306 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
and W
+
is the complex conjugate of the matrix W. Since W is a symmetric matrix, the
inverse matrix W
÷1
=
1
N
W
+
was used to derive (7.1.7b).
Example 7.3: Given x(n) = ¦1, 1, 0, 0¦, the DFT of this four-point sequence can be
computed using the matrix formulation as
X =
1 1 1 1
1 W
1
4
W
2
4
W
3
4
1 W
2
4
W
4
4
W
6
4
1 W
3
4
W
6
4
W
9
4

¸
¸
¸
¸
¸
¸
¸
¸
x
=
1 1 1 1
1 ÷j ÷1 j
1 ÷1 1 ÷1
1 j ÷1 ÷j

¸
¸
¸
¸
¸
¸
¸
¸
1
1
0
0

¸
¸
¸
¸
¸
¸
¸
¸
=
2
1 ÷j
0
1 ÷j

¸
¸
¸
¸
¸
¸
¸
¸
,
where we used symmetry and periodicity properties given in (7.1.5) to obtain
W
0
4
= W
4
4
= 1, W
1
4
= W
9
4
= ÷j, W
2
4
= W
6
4
= ÷1, and W
3
4
= j.
The IDFT can be computed with
x =
1
4
1 1 1 1
1 W
÷1
4
W
÷2
4
W
÷3
4
1 W
÷2
4
W
÷4
4
W
÷6
4
1 W
÷3
4
W
÷6
4
W
÷9
4

¸
¸
¸
¸
¸
X
=
1
4
1 1 1 1
1 j ÷1 ÷j
1 ÷1 1 ÷1
1 ÷j ÷1 j

¸
¸
¸
¸
2
1 ÷j
0
1 ÷j

¸
¸
¸
¸
=
1
1
0
0

¸
¸
¸
¸
:
As shown in Figure 7.1, the twiddle factors are equally spaced around the unit circle
at frequency intervals of f
s
=N (or 2p=N). Therefore the frequency samples X(k) repre-
sent discrete frequencies
f
k
= k
f
s
N
, k = 0, 1, . . . , N ÷1: (7:1:9)
The computational frequency resolution of the DFT is equal to the frequency increment
f
s
=N, and is sometimes referred to as the bin spacing of the DFT outputs. The spacing
DISCRETE FOURIER TRANSFORM 307
of the spectral lines depends on the number of data samples. This issue will be further
discussed in Section 7.3.2.
Since the DFT coefficient X(k) is a complex variable, it can be expressed in polar form
as
X(k) = [X(k)[e
j(k)
, (7:1:10)
where the DFT magnitude spectrum is defined as
[X(k)[ =

¦Re[X(k)[¦
2
÷¦Im[X(k)[¦
2

(7:1:11)
and the phase spectrum is defined as
(k) = tan
÷1
Im[X(k)[
Re[X(k)[

: (7:1:12)
These spectra provide a complementary way of representing the waveform, which
clearly reveals information about the frequency content of the waveform itself.
7.1.2 Important Properties of DFT
The DFT is important for the analysis of digital signals and the design of DSP systems.
Like the Fourier, Laplace, and z-transforms, the DFT has several important properties
that enhance its utility for analyzing finite-length signals. Many DFT properties are
similar to those of the Fourier transform and the z-transform. However, there are some
differences. For example, the shifts and convolutions pertaining to the DFT are circular.
Some important properties are summarized in this section. The circular convolution
property will be discussed in Section 7.1.3.
Linearity
If ¦x(n)¦ and ¦y(n)¦ are time sequences of the same length, then
DFT[ax(n) ÷by(n)[ = aDFT[x(n)[ ÷bDFT[y(n)[ = aX(k) ÷bY(k), (7:1:13)
where a and b are arbitrary constants. Linearity is a key property that allows us to
compute the DFTs of several different signals and determine the combined DFT via the
summation of the individual DFTs. For example, the frequency response of a given
system can be easily evaluated at each frequency component. The results can then be
combined to determine the overall frequency response.
308 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
Complex-conjugate property
If the sequence ¦x(n), 0 _ n _ N ÷1¦ is real, then
X(÷k) = X
+
(k) = X(N ÷k), 0 _ k _ N ÷1, (7:1:14)
where X
+
(k) is the complex conjugate of X(k). Or equivalently,
X(M ÷k) = X
+
(M ÷k), 0 _ k _ M, (7:1:15)
where M = N=2 if N is even, and M = (N ÷1)=2 if N is odd. This property shows that
only the first (M ÷1) DFT coefficients are independent. Only the frequency compon-
ents from k = 0 to k = M are needed in order to completely define the output. The rest
can be obtained from the complex conjugate of corresponding coefficients, as illustrated
in Figure 7.2.
The complex-conjugate (or symmetry) property shows that
Re[X(k)[ = Re[X(N ÷k)[, k = 1, 2, . . . , M ÷1 (7:1:16)
and
Im[X(k)[ = ÷Im[X(N ÷k)[, k = 1, 2, . . . , M ÷1: (7:1:17)
Thus the DFT of a real sequence produces symmetric real frequency components and
anti-symmetric imaginary frequency components about X(M). The real part of the DFT
output is an even function, and the imaginary part of the DFT output is an odd
function. From (7.1.16) and (7.1.17), we obtain
[X(k)[ = [X(N ÷k)[, k = 1, 2, . . . , M ÷1 (7:1:18)
and
f(k) = ÷f(N ÷k), k = 1, 2, . . . , M ÷1: (7:1:19)
Because of the symmetry of the magnitude spectrum and the anti-symmetry of the phase
spectrum, only the first M ÷1 outputs represent unique information from the input
signal. If the input to the DFT is a complex signal, however, all N complex outputs
could carry information.
X(0) X(1) … X(M−2) X(M−1) X(M) X(M+1) X(M+2) … X(N−1)
Complex conjugate
real real
Figure 7.2 Complex-conjugate property
DISCRETE FOURIER TRANSFORM 309
Periodicity
Because of the periodicity property shown in Figure 7.1, the DFT and IDFT produce
periodic results with period N. Therefore the frequency and time samples produced by
(7.1.3) and (7.1.6), respectively, are periodic with period N. That is,
X(k) = X(k ÷N) for all k (7:1:20a)
and
x(n) = x(n ÷N) for all n: (7:1:20b)
The finite-length sequence x(n) can be considered as one period of a periodic function
with period N. Also, the DFT X(k) is periodic with period N.
As discussed in Section 4.4, the spectrum of a discrete-time signal is periodic. For a
real-valued signal, the frequencies ranged from 0 to f
s
=2 were reversed for the range of 0
to ÷f
s
=2, and the entire range from ÷f
s
=2 to f
s
=2 was repeated infinitely in both
directions in the frequency domain. The DFT outputs represent a single period (from
0 to f
s
) of the spectrum.
Circular shifts
Let ¦X(k)¦ be the DFT of a given N-periodic sequence ¦x(n)¦, and let y(n) be a circular
shifted sequence defined by
y(n) = x(n ÷m)
mod N
, (7:1:21)
where m is the number of samples by which x(n) is shifted to the right (or delayed) and
the modulo operation
( j)
mod N
= j ÷
÷
iN (7:1:22a)
for some integer i such that
0 _ ( j)
mod N
< N: (7:1:22b)
For example, if m = 1, x(N ÷1) replaces x(0), x(0) replaces x(1), x(1) replaces x(2), etc.
Thus a circular shift of an N-point sequence is equivalent to a linear shift of its periodic
extension.
For a given y(n) in (7.1.21), we have
Y(k) = W
mk
N
X(k) = e
÷j(2pk=N)m
X(k): (7:1:23)
This equation states that the DFT coefficients of a circular-shifted N-periodic sequence
by m samples are a linear shift of X(k) by W
mk
N
.
310 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
DFT and z-transform
Consider a sequence x(n) having the z-transform X(z) with an ROC that includes
the unit circle. If X(z) is sampled at N equally spaced points on the unit circle at
z
k
= e
j2pk=N
, k = 0, 1, . . . , N ÷1, we obtain
X(z)[
z=e
j2pk=N =
¸
·
n=÷·
x(n)z
÷n

z=e
j2pk=N
=
¸
·
n=÷·
x(n)e
÷j(2pk=N)n
: (7:1:24)
This is identical to evaluating the discrete-time Fourier transform X(!) at the N equally
spaced frequencies !
k
= 2pk=N, k = 0, 1, . . . , N ÷1. If the sequence x(n) has a finite
duration of length N, the DFT of a sequence yields its z-transform on the unit circle at a
set of points that are 2p=N radians apart, i.e.,
X(k) = X(z)[
z = e
j
2p
N
( )
k
, k = 0, 1, . . . , N ÷1: (7:1:25)
Therefore the DFT is equal to the z-transform of a sequence x(n) of length N, evaluated
at N equally spaced points on the unit circle in the z-plane.
Example 7.4: Consider a finite-length DC signal
x(n) = c, n = 0, 1, . . . , N ÷1:
From (7.1.3), we obtain
X(k) = c
¸
N÷1
n=0
W
kn
N
= c
1 ÷W
kN
N
1 ÷W
k
N
:
Since W
kN
N
= e
÷j
2p
N
( )
kN
= 1 for all k and for W
k
N
= 1 for k = iN, we have X(k) = 0
for k = 1, 2, . . . , N ÷1. For k = 0,
¸
N÷1
n=0
W
kn
N
= N. Therefore we obtain
X(k) = cNd(k); k = 0, 1, . . . , N ÷1:
7.1.3 Circular Convolution
The Fourier transform, the Laplace transform, and the z-transform of the linear con-
volution of two time functions are simply the products of the transforms of the
individual functions. A similar result holds for the DFT, but instead of a linear
convolution of two sequences, we have a circular convolution. If x(n) and h(n) are
real-valued N-periodic sequences, y(n) is the circular convolution of x(n) and h(n)
defined as
y(n) = x(n) h(n), n = 0, 1, . . . , N ÷1, (7:1:26)
DISCRETE FOURIER TRANSFORM 311
x(n)
x(n−1)
x(n−2)
x(n−N+1)
h(n)
h(n−1)
h(n−2)
h(n−N+1)
Figure 7.3 Circular convolution of two sequences using the concentric circle approach
where denotes circular convolution. Then
Y(k) = X(k)H(k), k = 0, 1, . . . , N ÷1: (7:1:27)
Thus the circular convolution in time domain is equivalent to multiplication in the DFT
domain. Note that to compute the product defined in (7.1.27), the DFTs must be of
equal length. This means that the shorter of the two original sequences must be padded
with zeros to the length of the other before its DFT is computed.
The circular convolution of two periodic signals with period N can be expressed as
y(n) =
¸
N÷1
m=0
x(m)h(n ÷m)
mod N
=
¸
N÷1
m=0
h(m)x(n ÷m)
mod N
, (7:1:28)
where y(n) is also periodic with period N. This cyclic property of circular convolution
can be illustrated in Figure 7.3 by using two concentric rotating circles. To perform
circular convolution, N samples of x(n) [or h(n)] are equally spaced around the outer
circle in the clockwise direction, and N samples of h(n) [or x(n)] are displayed on the
inner circle in the counterclockwise direction starting at the same point. Corresponding
samples on the two circles are multiplied, and the resulting products are summed to
produce an output. The successive value of the circular convolution is obtained by
rotating the inner circle one sample in the clockwise direction; the result is computed by
summing the corresponding products. The process is repeated to obtain the next result
until the first sample of inner circle lines up with the first sample of the exterior circle
again.
Example 7.5: Given two 8-point sequences x(n) = ¦1, 1, 1, 1, 1, 0, 0, 0¦ and
h(n) = ¦0, 0, 0, 1, 1, 1, 1, 1¦. Using the circular convolution method illustrated in
Figure 7.3, we can obtain
312 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
n = 0, y(0) = 1 ÷1 ÷1 ÷1 = 4
1
0
1
1
1
1
1 1 0
0
0
0
1 1
1 0
n = 1, y(1) = 1 ÷1 ÷1 = 3
1
0
1
1
1
0
1 1 1
0
0
0
1 1
1 0
Repeating the process, we obtain
y(n) = x(n) h(n) = ¦4, 3, 2, 2, 2, 3, 4, 5¦:
This circular convolution is due to the periodicity of the DFT. In circular convolution,
the two sequences are always completely overlapping. As the end of one period is shifted
out, the beginning of the next is shifted in as shown in Figure 7.3. To eliminate the circular
effect and ensure that the DFTmethod results in a linear convolution, the signals must be
zero-padded so that the product terms from the end of the period being shifted out are
zero. Zero padding refers to the operation of extending a sequence of length N
1
to a length
N
2
(> N
1
) by appending (N
2
÷N
1
) zero samples to the tail of the given sequence. Note
that the padding number of zeros at the end of signal has no effect on its DTFT.
DISCRETE FOURIER TRANSFORM 313
Circular convolution can be used to implement linear convolution if both sequences
contain sufficient zero samples. The linear convolution of two sequences of lengths L
and M will result in a sequence of length L ÷M ÷1. Thus the two sequences must be
extended to the length of L ÷M ÷1 or greater by zero padding. That is, for the circular
convolution to yield the same result as the linear convolution, the sequence of length L
must be appended with at least M ÷1 zeros, while the sequence of length M must be
appended with at least L ÷1 zeros.
Example 7.6: Consider the previous example. If these 8-point sequences h(n) and
x(n) are zero-padded to 16 points, the resulting circular convolution is
y(n) = x(n) h(n) = ¦0, 0, 0, 1, 2, 3, 4, 5, 4, 3, 2, 1, 0, 0, 0, 0¦:
This result is identical to the linear convolution of the two sequences. Thus the
linear convolution discussed in Chapter 5 can be realized by the circular convolu-
tion with proper zero padding.
In MATLAB, zero padding can be implemented using the function zeros. For
example, the 8-point sequence x(n) given in example 7.5 can be zero-padded to 16
points with the following command:
x = [1, 1, 1, 1, zeros(1, 11)];
where the MATLAB function zeros(1, N)generates a row vector of N zeros.
7.2 Fast Fourier Transforms
The DFT is a very effective method for determining the frequency spectrum of a time-
domain signal. The only drawback with this technique is the amount of computation
necessary to calculate the DFT coefficients X(k). To compute each X(k), we need
approximately N complex multiplications and N complex additions based on the DFT
defined in (7.1.3). Since we need to compute N samples of X(k) for k = 0, 1, . . . , N ÷1,
a total of approximately N
2
complex multiplications and N
2
÷N complex additions are
required. When a complex multiplication is carried out using digital hardware, it
requires four real multiplications and two real additions. Therefore the number of
arithmetic operations required to compute the DFT is proportional to 4N
2
, which
becomes very large for a large number of N. In addition, computing and storing the
twiddle factors W
kn
N
becomes a formidable task for large values of N.
The same values of the twiddle factors W
kn
N
defined in (7.1.4) are calculated many
times during the computation of DFT since W
kn
N
is a periodic function with a limited
number of distinct values. Because W
N
N
= 1,
W
kn
N
= W
(kn)
mod N
N
for kn > N: (7:2:1)
For example, different powers of W
kn
N
have the same value as shown in Figure 7.1 for
N = 8. In addition, some twiddle factors have real or imaginary parts equal to 1 or 0. By
reducing these redundancies, a very efficient algorithm, called the FFT, exists. For
314 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
example, if N is a power of 2, then the FFT makes it possible to calculate the DFT with
N log
2
N operations instead of N
2
operations. For N = 1024, FFT requires about 10
4
operations instead 10
6
of operations for DFT.
The generic term FFT covers many different algorithms with different features,
advantages, and disadvantages. Each FFT has different strengths and makes different
tradeoffs between code complexity, memory usage, and computation requirements. The
FFT algorithm introduced by Cooley and Tukey requires approximately N log
2
N
multiplications, where N is a power of 2. The FFT can also be applied to cases where
N is a power of an integer other than 2. In this chapter, we introduce FFT algorithms
for the case where N is a power of 2, the radix-2 FFT algorithm.
There are two classes of FFT algorithms: decimation-in-time and decimation-in-
frequency. In the decimation-in-time algorithm, the input time sequence is successively
divided up into smaller sequences, and the DFTs of these subsequences are combined in
a certain pattern to yield the required DFT of the entire sequence with fewer operations.
Since this algorithm was derived by separating the time-domain sequence into succes-
sively smaller sets, the resulting algorithm is referred to as a decimation-in-time algo-
rithm. In the decimation-in-frequency algorithm, the frequency samples of the DFT are
decomposed into smaller and smaller subsequences in a similar manner.
7.2.1 Decimation-in-Time
In the decimation-in-time algorithm, the N-sample sequence ¦x(n), n = 0, 1, . . . , N ÷1¦
is first divided into two shorter interwoven sequences: the even numbered sequence
x
1
(m) = x(2m), m = 0, 1, . . . , (N=2) ÷1 (7:2:2a)
and the odd numbered sequence
x
2
(m) = x(2m÷1), m = 0, 1, . . . , (N=2) ÷1: (7:2:2b)
The DFT expressed in (7.1.3) can be divided into two DFTs of length N=2. That is,
X(k) =
¸
N÷1
n=0
x(n)W
kn
N
=
¸
(N=2)÷1
m=0
x(2m)W
2mk
N
÷
¸
(N=2)÷1
m=0
x(2m÷1)W
(2m÷1)k
N
: (7:2:3)
Since
W
2mk
N
= e
÷j
2p
N
2mk
= e
÷j
2p
N=2
mk
= W
mk
N=2
, (7:2:4)
Equation (7.2.3) can be written as
X(k) =
¸
(N=2)÷1
m=0
x
1
(m)W
mk
N=2
÷W
k
N
¸
(N=2)÷1
m=0
x
2
(m)W
mk
N=2
, (7:2:5)
FAST FOURIER TRANSFORMS 315
where each of the summation terms is reduced to an N=2 point DFT. Furthermore,
from symmetry and periodicity properties given in (7.1.5), Equation (7.2.5) can be
written as
X(k) = X
1
(k) ÷W
k
N
X
2
(k); k = 0, 1, . . . , N ÷1
=
X
1
(k) ÷W
k
N
X
2
(k), k = 0, 1, . . . , (N=2) ÷1
X
1
(k) ÷W
k
N
X
2
(k), k = N=2, . . . , N ÷1,

(7:2:6)
where X
1
(k) = DFT[x
1
(m)[ and X
2
(k) = DFT[x
2
(m)[ using the N=2-point DFT.
The important point about this result is that the DFT of N samples becomes a linear
combination of two smaller DFTs, each of N=2 samples. This procedure is illustrated in
Figure 7.4 for the case N = 8. The computation of X
1
(k) and X
2
(k) requires 2(N=2)
2
multiplications, the computation of W
k
N
X
2
(k) requires N=2 multiplications. This gives a
total of approximately (N
2
÷N)=2 multiplications. Compared with N
2
operations for
direct evaluation of the DFT, there is a saving in computation when N is large after only
one stage of splitting signals into even and odd sequences. If we continue with this
process, we can break up the single N-point DFT into log
2
N DFTs of length 2. The
final algorithm requires computation proportional to N log
2
N, a significant saving over
the original N
2
.
Equation (7.2.6) is commonly referred to as the butterfly computation because of its
crisscross appearance, which can be generalized in Figure 7.5. The upper group gen-
erates the upper half of the DFT coefficient vector X, and the lower group generates the
−1
−1
−1
−1
x(0)
x(2)
x(4)
x(6)
x(1)
x(3)
x(5)
x(7)
N/2-point
DFT
N/2-point
DFT
X
1
(0)
X
1
(1)
X
1
(2)
X
1
(3)
X
2
(0)
X
2
(1)
X
2
(2)
X
2
(3)
X(0)
X(1)
X(2)
X(3)
X(4)
X(5)
X(6)
X(7)
W
8
0
W
8
1
W
8
2
W
8
3
Figure 7.4 Decomposition of an N-point DFT into two N=2 DFTs, N = 8
W
N
k
−1
(m−1)th
stage
mth
stage
Figure 7.5 Flow graph for a butterfly computation
316 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
lower half. Each butterfly involves just a single complex multiplication by a twiddle
factor W
k
N
, one addition, and one subtraction. For this first decomposition, the twiddle
factors are indexed consecutively, and the butterfly values are separated by N/2 samples.
The order of the input samples has also been rearranged (split between even and odd
numbers), which will be discussed in detail later.
Since N is a power of 2, N=2 is even. Each of these N=2-point DFTs in (7.2.6) can be
computed via two smaller N=4-point DFTs, and so on. The second step process is
illustrated in Figure 7.6. Note that the order of the input samples has been rearranged
as x(0), x(4), x(2), and x(6) because x(0), x(2), x(4), and x(6) are considered to be the 0th,
1st, 2nd, and 3rd inputs in a 4-point DFT. Similarly, the order of x(1), x(5), x(3), and x(7)
is used in the second 4-point DFT.
By repeating the process associated with (7.2.6), we will finally end up with a set of 2-
point DFTs since N is a power of 2. For example, in Figure 7.6, the N/4-point DFT
became a 2-point DFT since N = 8. Since the twiddle factor for the first stage, W
0
N
= 1,
the 2-point DFT requires only one addition and one subtraction. The 2-point DFT
illustrated in Figure 7.7 is identical to the butterfly network.
Example 7.7: Consider the two-point DFT algorithm which has two input time-
domain samples x(0) and x(1). The output frequency-domain samples are X(0)
and X(1). For this case, the DFT can be expressed as
X(k) =
¸
1
n=0
x(n)W
nk
2
, k = 0, 1:
−1
−1
−1
−1
−1
−1
−1
−1
x(0)
x(4)
x(2)
x(6)
x(1)
x(5)
x(3)
x(7)
N/4-point
DFT
N/4-point
DFT
N/4-point
DFT
N/4-point
DFT
X(0)
X(1)
X(2)
X(3)
X(4)
X(5)
X(6)
X(7)
W
8
0
W
8
1
W
8
2
W
8
3
W
8
0
W
8
2
W
8
0
W
8
2
Figure 7.6 Flow graph illustrating second step of N-point DFT, N = 8
−1
Figure 7.7 Flow graph of 2-point DFT
FAST FOURIER TRANSFORMS 317
Since W
0
2
= 1 and W
1
2
= e
÷p
= ÷1, we have
X(0) = x(0) ÷x(1)
and
X(1) = x(0) ÷x(1):
The signal flow graph is shown in Figure 7.7. Note that the results are agreed with
the results obtained in Example 7.1.
As shown in Figure 7.6, the output sequence is in natural order, while the input
sequence has the unusual order. Actually the order of the input sequence is arranged as
if each index was written in binary form and then the order of binary digits was reversed.
The bit-reversal process is illustrated in Table 7.1 for the case N = 8. Each of the time
sample indices in decimal is converted to its binary representation. The binary bit streams
are then reversed. Converting the reversed binary numbers to decimal values gives the
reordered time indices. If the input is in natural order the output will be in bit-reversed
order. We can either shuffle the input sequence with a bit-reversal algorithm to get the
output sequence in natural order, or let the input sequence be in natural order and shuffle
the bit-reversed results to obtain the output in natural order. Note that most modern
DSP chips such as the TMS320C55x provide the bit-reversal addressing mode to support
this bit-reversal process. Therefore the input sequence can be stored in memory with the
bit-reversed addresses computed by the hardware.
For the FFT algorithm shown in Figure 7.6, once all the values for a particular stage
are computed, the old values that were used to obtain them are never required again.
Thus the FFT needs to store only the N complex values. The memory locations used for
the FFT outputs are the same as the memory locations used for storing the input data.
This observation is used to produce in-place FFT algorithms that use the same memory
locations for input samples, all intermediate calculations, and final output numbers.
Table 7.1 Example of bit-reversal process, N = 8 (3-bit)
Input sample index
&
Bit-reversed sample index
Decimal Binary Binary Decimal
0 000 000 0
1 001 100 4
2 010 010 2
3 011 110 6
4 100 001 1
5 101 101 5
6 110 011 3
7 111 111 7
318 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
7.2.2 Decimation-in-Frequency
The development of the decimation-in-frequency FFT algorithm is similar to the
decimation-in-time algorithm presented in the previous section. The first step consists
of dividing the data sequence into two halves of N/2 samples. Then X(k) in (7.1.3) can be
expressed as the sum of two components to obtain
X(k) =
¸
(N=2)÷1
n=0
x(n)W
nk
N
÷
¸
N÷1
n=N=2
x n ( )W
nk
N
=
¸
(N=2)÷1
n=0
x(n)W
nk
N
÷W
(N=2)k
N
¸
(N=2)÷1
n=0
x n ÷
N
2

W
nk
N
: (7:2:7)
Using the fact that
W
N=2
N
= e
÷j
2p
N
(N=2)
= e
÷jp
= ÷1, (7:2:8)
Equation (7.2.7) can be simplified to
X(k) =
¸
(N=2)÷1
n=0
x(n) ÷(÷1)
k
x n ÷
N
2
¸
W
nk
N
: (7:2:9)
The next step is to separate the frequency terms X(k) into even and odd samples of k.
Since W
2kn
N
= W
kn
N=2
, Equation (7.2.9) can be written as
X(2k) =
¸
(N=2)÷1
n=0
x(n) ÷x n ÷
N
2
¸
W
kn
N=2
(7:2:10a)
and
X(2k ÷1) =
¸
(N=2)÷1
n=0
x(n) ÷x n ÷
N
2
¸
W
k
N
W
kn
N=2
(7:2:10b)
for 0 _ k _ (N=2) ÷1. Let x
1
(n) = x(n) ÷x n ÷
N
2

and x
2
(n) = x(n) ÷x n ÷
N
2

for
0 _ n _ (N=2) ÷1, the first decomposition of an N-point DFT into two N=2-point
DFTs is illustrated in Figure 7.8.
Again, the process of decomposition is continued until the last stage is made up of
two-point DFTs. The decomposition proceeds from left to right for the decimation-
in-frequency development and the symmetry relationships are reversed from the
decimation-in-time algorithm. Note that the bit reversal occurs at the output instead
of the input and the order of the output samples X(k) will be re-arranged as bit-
reversed samples index given in Table 7.1. The butterfly representation for the
decimation-in-frequency FFT algorithm is illustrated in Figure 7.9.
FAST FOURIER TRANSFORMS 319
−1
−1
−1
−1
x(0)
x(1)
x(2)
x(3)
x(4)
x(5)
x(6)
x(7)
N/2-point
DFT
N/2-point
DFT
x
1
(0)
x
1
(1)
x
1
(2)
x
1
(3)
x
2
(0)
x
2
(1)
x
2
(2)
x
2
(3)
X(0)
X(2)
X(4)
X(6)
X(1)
X(3)
X(5)
X(7)
W
8
0
W
8
1
W
8
2
W
8
3
Figure 7.8 Decomposition of an N-point DFT into two N=2 DFTs using decimation-in-
frequency algorithm, N = 8
−1
W
N
k
(m−1)th
stage
mth
stage
Figure 7.9 Butterfly network for decimation-in-frequency FFT algorithm
7.2.3 Inverse Fast Fourier Transform
The FFT algorithms introduced in the previous section can be easily modified to
compute the IFFT efficiently. This is apparent from the similarities of the DFT and
the IDFT definitions given in (7.1.3) and (7.1.6), respectively. Complex conjugating
(7.1.6), we have
x
+
(n) =
1
N
¸
N÷1
k=0
X
+
(k)W
kn
N
, n = 0, 1, . . . , N ÷1: (7:2:11)
Therefore an FFT algorithm can be used to compute the inverse DFT by first con-
jugating the DFT coefficients X(k) to obtain X
+
(k), computing the DFT of X
+
(k) use an
FFT algorithm, scaling the results by 1/N to obtain x
+
(n), and then complex conjugating
x
+
(n) to obtain the output sequence x(n). If the signal is real-valued, the final conjuga-
tion operation is not required.
All the FFT algorithms introduced in this chapter are based on two-input, two-
output butterfly computations, and are classified as radix-2 complex FFT algorithms.
It is possible to use other radix values to develop FFT algorithms. However, these
algorithms do not work well when the length is a number with few factors. In addition,
these algorithms are more complicated than the radix-2 FFT algorithms, and the
320 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
routines are not as available for DSP processors. Radix-2 and radix-4 FFT algorithms
are the most common, although other radix values can be employed. Different radix
butterflies can be combined to form mixed-radix FFT algorithms.
7.2.4 MATLAB Implementations
As introduced in Section 4.4.4, MATLAB provides a function
y = fft(x);
to compute the DFT of time sequence x(n) in the vector x. If x is a matrix, y is the DFT
of each column of the matrix. If the length of the x is a power of 2, the fft function
employs a high-speed radix-2 FFT algorithm. Otherwise a slower mixed-radix algo-
rithm is employed.
An alternative way of using fft function is
y = fft(x, N);
to perform N-point FFT. If the length of x is less than N, then the vector x is
padded with trailing zeros to length N. If the length of x is greater than N, fft
function truncates the sequence x and only performs the FFT of the first N samples of
data.
The execution time of the fft function depends on the input data type and the
sequence length. If the input data is real-valued, it computes a real power-of-two FFT
algorithm that is faster than a complex FFT of the same length. As mentioned earlier,
the execution is fastest if the sequence length is exactly a power of 2. For this reason, it
is usually better to use power-of-two FFT. For example, if the length of x is 511,
the function y = fft(x, 512)will be computed faster than fft(x), which performs
511-point DFT.
It is important to note that the vectors in MATLAB are indexed from 1 to N instead
of from 0 to N ÷1 given in the DFT and the IDFT definitions. Therefore the relation-
ship between the actual frequency in Hz and the frequency index k given in (7.1.9) is
modified as
f
k
= (k ÷1)
f
s
N
, k = 1, 2, . . . , N (7:2:12)
for indexing into the y vector that contains X(k).
The IFFT algorithm is implemented in the MATLAB function ifft, which can be
used as
y = ifft(x);
or
y = ifft(x,N);
The characteristics and usage of ifft are the same as those for fft.
FAST FOURIER TRANSFORMS 321
7.3 Applications
FFT has a wide variety of applications in DSP. Spectral analysis often requires the
numerical computation of the frequency spectrum for a given signal with large sample
sets. The DFT is also used in the coding of waveforms for efficient transmission or
storage. In these cases the FFT may provide the only possible means for spectral
computation within the limits of time and computing cost. In this section, we will also
show how the FFT can be used to implement linear convolution for FIR filtering in a
computationally efficient manner.
7.3.1 Spectrum Estimation and Analysis
The spectrum estimation techniques may be categorized as non-parametric and para-
metric. The non-parametric methods that include the periodogram have the advantage
of using the FFT for efficient implementation, but have the disadvantage of having
limited frequency resolution for short data lengths. Parametric methods can provide
higher resolution. The most common parametric technique is to derive the spectrum
from the parameters of an autoregressive model of the signal. In this section, we will
introduce the principles and practice of the spectral estimation and analysis using non-
parametric methods.
The inherent properties of the DFT directly relate to the performance of spectral
analysis. If a discrete-time signal is periodic, it is possible to calculate its spectrum using
the DFT. For an aperiodic signal, we can break up long sequence into smaller segments
and analyze each individual segment using the FFT. This is reasonable because the very
long signal probably consists of short segments where the spectral content does not
change. The spectrum of a signal of length L can be computed using an N-point FFT. If
L < N, we must increase the signal length from L to N by appending N ÷L zero
samples to the tail of the signal.
To compute the spectrum of an analog signal digitally, the signal is sampled first and
then transformed to the frequency domain by a DFT or an FFT algorithm. As discussed
in Chapter 3, the sampling rate f
s
must be high enough to minimize aliasing effects. As
discussed in Chapter 4, the spectrum of the sampled signal is the replication of the
desired analog spectrum at multiples of the sampling frequency. The proper choice of
sampling rate can guarantee that these two spectra are the same over the Nyquist
interval. The DFT of an arbitrary set of sampled data may not be always the true
DFT of the signal from which the data was obtained. This is because the signal is
continuous, whereas the data set is truncated at its beginning and end. The spectrum
estimated from a finite number of samples is correct only if the signal is periodic and the
sample set has exactly one period. In practice, we may not have exactly one period of the
periodic signal as the input.
As discussed in Section 7.1, the computational frequency resolution of the N-point
DFT is f
s
=N. The DFT coefficients X(k) represent frequency components equally
spaced at frequencies
f
k
=
kf
s
N
, k = 0, 1, . . . , N ÷1: (7:3:1)
322 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
If there is a signal component that falls between two adjacent frequency components in
the spectrum, it cannot be properly represented. Its energy will be shared between
neighboring bins and the nearby spectral amplitude will be distorted.
Example 7.8: Consider a sinewave of frequency f = 30 Hz expressed as
x(n) = sin(2pfnT), n = 0, 1, . . . , 127,
where the sampling period is T = 1=128 seconds. Because the computational
frequency resolution ( f
s
=N) is 1 Hz using a 128-point FFT, the line component
at 30 Hz can be represented by X(k) at k = 30. The amplitude spectrum is shown in
Figure 7.10a. The only non-zero term that occurs in [X(k)[ is at k = 30, which
corresponds to 30 Hz. It is clear that a magnitude spectrum with a single non-zero
spectral component is what one would expect for a pure sinusoid.
Example 7.9: Consider a different sinewave of frequency 30.5 Hz with sampling
rate 128 Hz. The magnitude spectrum in Figure 7.10b shows several spectral
components that are symmetric about 30.5 Hz. From evaluating the DFT coeffi-
cients, we cannot tell the exact frequency of the sinusoid. The reason why [X(k)[
does not have a single spectral component is that the line component at 30.5 Hz is
in between k = 30 and k = 31. The frequency components in x(n) that are not
equal to f
k
given in (7.3.1) tend to spread into portions of the spectrum that are
adjacent to the correct spectral value. The MATLAB program (Fig7_10.m in the
software package) for Examples 7.8 and 7.9 is listed as follows:
10 20
d
B
30 40 50 60 0
60
40
20
0
10 20
d
B
(a)
(b)
30
Frequency, Hz
40 50 60 0
60
40
20
0
Figure 7.10 Effect of computational frequency resolution on sinewave spectra: (a) sinewave at
30 Hz, and (b) sinewave at 30.5 Hz
APPLICATIONS 323
n = [0:127];
x1 = sin(2*pi*30*n/128); X1 = abs(fft(x1));
x2 = sin(2*pi*30.5*n/128); X2 = abs(fft(x2));
subplot(2,1,1),plot(n,X1), axis( [0 64 0 70]),
title(`(a) Sinewave at 30Hz'),
subplot(2,1,2),plot(n,X2), axis([0 64 0 70]),
title(`(b) Sinewave at 30.5Hz'),
xlabel(`Frequency,Hz');
A solution to this problem is to make the frequencies f
k
= kf
s
=N more closely spaced,
thus matching the signal frequencies. This may be achieved by using a larger DFT size
N to increase the computational frequency resolution of the spectrum. If the number
of data samples is not sufficiently large, the sequence may be expanded by adding
additional zeros to the true data, thus increasing the length N. The added zeros serve
to increase the computational frequency resolution of the estimated spectrum to the
true spectrum without adding additional information. This process is simply the
interpolation of the spectral curve between adjacent frequency components. A real
improvement in frequency resolution can only be achieved if a longer data record is
available.
A number of problems have to be avoided in performing non-parametric spectral
analysis such as aliasing, finite data length, spectral leakage, and spectral smearing.
The effects of spectral leakage and smearing may be minimized by windowing the
data using a suitable window function. These issues will be discussed in the following
section.
7.3.2 Spectral Leakage and Resolution
The data that represents the signal of length N is effectively obtained by multiplying all
the sampled values in the interval by one, while all values outside this interval are
multiplied by zero. This is equivalent to multiplying the signal by a rectangular window
of width N and height 1, expressed as
w(n) =
1, 0 _ n _ N ÷1
0, otherwise.

(7:3:2)
In this case, the sampled data x
N
(n) is obtained by multiplying the signal x(n) with the
window function w(n). That is,
x
N
(n) = w(n)x(n) =
x(n), 0 _ n _ N ÷1
0, otherwise.

(7:3:3)
The multiplication of x(n) by w(n) ensures that x
N
(n) vanishes outside the window. As
the length of the window increases, the windowed signal x
N
(n) becomes a better
approximation of x(n), and thus X(k) becomes a better approximation of the DTFT
X(!).
324 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
The time-domain multiplication given in (7.3.3) is equivalent to the convolution in the
frequency domain. Thus the DFT of x
N
(n) can be expressed as
X
N
(k) = W(k) + X(k) =
¸
N
l=÷N
W(k ÷l)X(k), (7:3:4)
where W(k) is the DFTof the windowfunctionw(n), and X(k) is the true DFTof the signal
x(n). Equation (7.3.4) shows that the computed spectrum consists of the true spectrum
X(k) convoluted with the window function's spectrum W(k). This means that when we
apply a window to a signal, the frequency components of the signal will be corrupted in
the frequency domain by a shifted and scaled version of the window's spectrum.
As discussed in Section 5.3.3, the magnitude response of the rectangular window
defined in (7.3.2) can be expressed as

W(!)

=
sin(!N=2)
sin(!=2)

: (7:3:5)
It consists of a mainlobe of height N at ! = 0, and several smaller sidelobes. The
frequency components that lie under the sidelobes represent the sharp transition of
w(n) at the endpoints. The sidelobes are between the zeros of W(!), with center
frequencies at ! =
(2k÷1)p
N
, k =
÷
÷
1,
÷
÷
2, . . ., and the first sidelobe is down only 13 dB
from the mainlobe level. As N increases, the height of the mainlobe increases and its
width becomes narrower. However, the peak of the sidelobes also increases. Thus the
ratio of the mainlobe to the first sidelobe remains the same about 13 dB.
The sidelobes introduce spurious peaks into the computed spectrum, or to cancel true
peaks in the original spectrum. The phenomenon is known as spectral leakage. To avoid
spectral leakage, it is necessary to use a shaped window to reduce the sidelobe effects.
Suitable windows have a value of 1 at n = M, and are tapered to 0 at points n = 0 and
n = N ÷1 to smooth out both ends of the input samples to the DFT.
If the signal x(n) consists of a single sinusoid, that is,
x(n) = cos(!
0
n), (7:3:6)
the spectrum of the infinite-length sampled signal over the Nyquist interval is given as
X(!) = 2pd(! ÷
÷
!
0
), ÷p _ ! _ p, (7:3:7)
which consists of two line components at frequencies ÷
÷
!
0
. However, the spectrum of the
windowed sinusoid defined in (7.3.3) can be obtained as
X
N
(!) =
1
2
[W(! ÷!
0
) ÷W(! ÷!
0
)[, (7:3:8)
where W(!) is the spectrum of the window function.
Equation (7.3.8) shows that the windowing process has the effect of smearing the
original sharp spectral line d(! ÷!
0
) at frequency !
0
and replacing it with W(! ÷!
0
).
APPLICATIONS 325
Thus the power of the infinite-length signal that was concentrated at a single frequency
has been spread into the entire frequency range by the windowing operation. This
undesired effect is called spectral smearing. Thus windowing not only distorted the
spectrum due to leakage effects, but also reduced spectral resolution. For example, a
similar analysis can be made in the case when the signal consists of two sinusoidal
components. That is,
x(n) = cos(!
1
n) ÷cos(!
2
n): (7:3:9)
The spectrum of the windowed signal is
X
N
(!) =
1
2
[W(! ÷!
1
) ÷W(! ÷!
1
) ÷W(! ÷!
2
) ÷W(! ÷!
2
)[: (7:3:10)
Again, the sharp spectral lines are replaced with their smeared versions. From (7.3.5),
the spectrum W(!) has its first zero at frequency ! = 2p=N. If the frequency separation,
D! = [!
1
÷!
2
[, of the two sinusoids is
D! _
2p
N
, (7:3:11)
the mainlobe of the two window functions W(! ÷!
1
) and W(! ÷!
2
) overlap. Thus the
two spectral lines in X
N
(!) are not distinguishable. This undesired effect starts when D!
is approximately equal to the mainlobe width 2p=N. Therefore the frequency resolution
of the windowed spectrum is limited by the window's mainlobe width.
To guarantee that two sinusoids appear as two distinct ones, their frequency separ-
ation must satisfy the condition
D! >
2p
N
, (7:3:12a)
in radians per sample, or
Df >
f
s
N
, (7:3:12b)
in Hz. Thus the minimum DFT length to achieve a desired frequency resolution is given
as
N >
f
s
Df
=
2p
D!
: (7:3:13)
In summary, the mainlobe width determines the frequency resolution of the windowed
spectrum. The sidelobes determine the amount of undesired frequency leakage. The
optimum window used for spectral analysis must have narrow mainlobe and small
sidelobes. Although adding to the record length by zero padding increases the FFT size
and thereby results in a smaller Df , one must be cautious to have sufficient record length
to support this resolution.
326 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
In Section 5.3.4, we used windows to smooth out the truncated impulse response of
an ideal filter for designing an FIR filter. In this section, we showed that those window
functions can also be used to modify the spectrum estimated by the DFT. If the window
function, w(n), is applied to the input signal, the DFT outputs are given by
X(k) =
¸
N÷1
n=0
w(n)x(n)W
kn
N
, k = 0, 1, . . . , N ÷1: (7:3:14)
The rectangular window has the narrowest mainlobe width, thus providing the best
spectral resolution. However, its high-level sidelobes produce undesired spectral leak-
age. The amount of leakage can be substantially reduced at the cost of decreased
spectral resolution by using appropriate non-rectangular window functions introduced
in Section 5.3.4.
As discussed before, frequency resolution is directly related to the window's mainlobe
width. A narrow mainlobe will allow closely spaced frequency components to be
identified; while a wide mainlobe will cause nearby frequency components to blend.
For a given window length N, windows such as rectangular, Hanning, and Hamming
have relatively narrow mainlobe compared with Blackman or Kaiser windows. Unfor-
tunately, the first three windows have relatively high sidelobes, thus having more
spectral leakage. There is a trade-off between frequency resolution and spectral leakage
in choosing windows for a given application.
Example 7.10: Consider the sinewave used in Example 7.9. Using the Kaiser
window defined in (5.3.26) with L = 128 and b = 8:96, the magnitude spectrum
is shown in Figure 7.11 using the MATLAB script Exam7_10.m included in the
software package.
10 20
d
B
30 40 50 60 0
60
40
20
0
10 20
d
B
(a)
(b)
30
Frequency, Hz
40 50 60 0
60
40
20
0
Figure 7.11 Effect of Kaiser window function for reducing spectral leakage: (a) rectangular
window, and (b) Kaiser window
APPLICATIONS 327
An effective method for decreasing the mainlobe width is by increasing the window
length. For a given window, increasing the length of the window reduces the width of
the mainlobe, which leads to better frequency resolution. However, if the signal changes
frequency content over time, the window cannot be too long in order to provide a
meaningful spectrum. In addition, a longer window implies using more data, so there is
a trade-off between frequency resolution and the cost of implementation. If the number
of available signal samples is less than the required length, we can use the zero padding
technique. Note that the zeros are appended after windowing is performed.
7.3.3 Power Density Spectrum
The finite-energy signals possess a Fourier transform and are characterized in the
frequency domain by their power density spectrum. Consider a sequence x(n) of length
N whose DFT is X(k), Parseval's theorem can be expressed as
E =
¸
N÷1
n=0

x(n)

2
=
1
N
¸
N÷1
k=0

X(k)

2
: (7:3:15)
The term

X(k)

2
is called the power spectrum and is a measure of power in signal at
frequency f
k
defined in (7.3.1). The DFT magnitude spectrum [X(k)[ is defined in
(7.1.11). Squaring the magnitude of the DFT coefficient produces a power spectrum,
which is also called the periodogram.
As discussed in Section 3.3, stationary random processes do not have finite energy
and thus do not possess Fourier transform. Such signals have a finite average power and
are characterized by the power density spectrum (PDS) defined as
P(k) =
1
N

X(k)

2
=
1
N
X(k)X
+
(k), (7:3:16)
which is also commonly referred to as the power spectral density, or simply power
spectrum.
The PDS is a very useful concept in the analysis of random signals since it provides a
meaningful measure for the distribution of the average power in such signals. There are
many different techniques developed for estimating the PDS. Since the periodogram is
not a consistent estimate of the true PDS, the periodogram averaging method may be
used to reduce statistical variation of the computed spectra. Given a signal vector xn
which consists N samples of digital signal x(n), a crude estimate of the PDS using
MATLAB is
pxn = abs(fft(xn, 1024)).^2/N;
In practice, we only have a finite-length sequence whose PDS is desired. One way of
computing the PDS is to decompose x(n) into M segments, x
m
(n), of N samples each.
These signal segments are spaced N/2 samples apart, i.e., there is 50 percent overlap
between successive segments as illustrated in Figure 7.12.
328 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
m = 0
m = 1
m = 2
N
N/2
N/2
m = M−1
N
x
0
(n)
x
1
(n)
x
2
(n)
x
M−1
(n)
Figure 7.12 Segments used to estimate power density spectrum
In order to reduce spectral leakage, each x
m
(n) is multiplied by a non-rectangular
window function w(n) of length N. This results in `windowed segments' x
/
m
(n), which are
given by
x
/
m
(n) = w(n)x
m
(n); n = 0, . . . , N ÷1, (7:3:17)
for 0 _ m _ M ÷1. The windowing operation results in reduction of frequency resolu-
tion, which may be compensated by increasing the length N.
We then compute the DFT of x
/
m
(n) given in (7.3.17) to get X
/
m
(k). The mth period-
ogram is defined as
P
m
(k) =
1
NP
w

X
/
m
(k)

2
, 0 _ k _ N ÷1, (7:3:18)
where
P
w
=
1
N
¸
N÷1
n=0
w
2
(n) (7:3:19)
is a normalization factor for the average power in the window sequence w(n). Finally,
the desired PDS is the average of these periodograms. That is,
P(k) =
1
M
¸
M÷1
m=0
P
m
(k) =
1
MNP
w
¸
M÷1
m=0

X
/
m
(k)

2
, 0 _ k _ N ÷1: (7:3:20)
Therefore the PDS estimate given in (7.3.20) is a weighted sum of the periodograms of
each of the individual overlapped segments. The 50 percent overlap between successive
segments helps to improve certain statistical properties of this estimation.
The Signal Processing Toolbox provides the function psd to average the period-
ograms of windowed segments of a signal. This MATLAB function estimates the
PDS of the signal given in the vector x using the following statement:
pxx = psd(x, nfft, Fs, window, noverlap);
APPLICATIONS 329
where nfft specifies the FFT length, Fs is the sampling frequency, window specifies
the selected window function, and noverlap is the number of samples by which the
segments overlap.
7.3.4 Fast Convolution
As discussed in Chapter 5, FIR filtering is a linear convolution of the finite impulse
response h(n) with the input sequence x(n). If the FIR filter has L coefficients, we need L
real multiplications and L ÷1 real additions to compute each output y(n). To obtain L
output samples, the number of operations (multiplication and addition) needed is
proportional to L
2
. To reduce the computational requirements, we consider the alter-
nate approach defined in (7.1.26) and (7.1.27), which uses the property that the con-
volution of the sequences x(n) and h(n) is equivalent to multiplying their DFTs X(k)
and H(k).
As described in Section 7.1.3, the linear convolution can be implemented using zero
padding if the data sequence x(n) also has finite duration. To take advantage of efficient
FFT and IFFT algorithms, we use these computational efficient algorithms as illus-
trated in Figure 7.13. The procedure of using FFT to implement linear convolution in a
computationally efficient manner is called the fast convolution. Compared to the direct
implementation of FIR filtering, fast convolution will provide a significant reduction in
computational requirements for higher order FIR filters, thus it is often used to imple-
ment FIR filtering in applications having long data samples.
It is important to note that the fast convolution shown in Figure 7.13 produces the
circular convolution discussed in Section 7.1.3. In order to produce a filter result equal
to a linear convolution, it is necessary to append zeros to the signals in order to
overcome the circular effect. If the data sequence x(n) has finite duration M, the first
step is to pad both sequences with zeros to a length corresponding to an allowable FFT
size N (_ L ÷M ÷1), where L is the length of the sequence h(n). The FFT is computed
for both sequences, the complex products defined in (7.1.27) are calculated, and the
IFFT is used to obtain the results. The desired linear convolution is contained in the
first L ÷M ÷1 terms of these results.
The FFT of the zero-padded data requires about N log
2
N complex computations.
Since the filter impulse response h(n) is known as a priori, the FFT of the impulse
response can be pre-calculated and stored. The computation of product
Y(k) = H(k)X(k) takes N complex multiplications, and the inverse FFT of the pro-
ducts Y(k) requires another N log
2
N complex multiplications. Therefore the fast con-
volution shown in Figure 7.13 requires about (8N log
2
N ÷4N) real multiplications.
x(n)
h(n)
X(k)
H(k)
Y(k) y(n)
FFT
FFT
IFFT
Figure 7.13 Fast convolution algorithm
330 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
Compared with LM required by direct FIR filtering, the computational saving is
significant when both L and M are large.
For many applications, the input sequence is very long compared to the order of FIR
filter L. This is especially true in real-time applications where the input sequence is of
infinite duration. In order to use the efficient FFT and IFFT algorithms, the input
sequence must be grouped into segments of N (N > L and N is a size supported by the
FFT algorithm) samples, process each segment using the FFT, and finally assemble the
output sequence from the outputs of each segment. This procedure is called the block
processing operation. The selection of the block size is a function of the filter order.
There will be end effects from each convolution that must be accounted for as the
segments are recombined to produce the output sequence. There are two techniques for
the segmentation and recombination of the data: the overlap-save and overlap-add
algorithms.
Overlap-save technique
The overlap-save process is carried out by overlapping L input samples on each
segment, where L is the order of the FIR filter. The output segments are truncated to
be non-overlapping and then concatenated. The process is illustrated in Figure 7.14 and
is described by the following steps:
1. Perform N-point FFT of the expanded (zero padded) impulse response sequence
h
/
(n) =
h(n), n = 0, 1, . . . , L ÷1
0, n = L, L ÷1, . . . , N ÷1 ,

(7:3:21)
2. to obtain H
/
(k) k = 0, 1, . . . , N ÷1, where h(n) is the impulse response of the FIR
filter. Note that for a time-invariant filter, this process can be pre-calculated off-line
and stored in memory.
2. Select N signal samples x
m
(n) (where m is the segment index) from the input
sequence x(n) based on the overlap illustrated in Figure 7.14, and then use N-point
m−1
0 N−1
m
m+1
L
x
m−1
(n)
x
m+1
(n)
x
m
(n)
y
m−1
(n)
y
m+1
(n)
y
m
(n)
L
discarded
m+1
m
m−1
Figure 7.14 Overlap data segments for the overlap-save technique
APPLICATIONS 331
2. FFT to obtain X
m
(k). Clearly, to avoid excessive overlap, we usually choose N L
and N is the size supported by the FFT algorithms such as power of 2 for using
radix-2 algorithms.
3. Multiply the stored H
/
(k) (obtained in step 1) by the X
m
(k) of segment m (obtained
in step 2) to get
Y
m
(k) = H
/
(k)X
m
(k), k = 0, 1, . . . , N ÷1: (7:3:22)
4. Perform N-point IFFT of Y
m
(k) to obtain y
m
(n), n = 0, 1, . . . , N ÷1.
5. Discard the first L samples from each successive IFFT output since they are
circularly wrapped and superimposed as discussed in Section 7.1.3. The resulting
segments of (N ÷L) samples are concatenated to produce y(n).
Overlap-add technique
In the overlap-add process, the input sequence x(n) is divided into non-overlapping
segments of length (N ÷L). Each segment is zero-padded to produce x
m
(n) of length N.
Following the steps 2, 3, and 4 of the overlap-save method to obtain N-point segments
y
m
(n). Since the convolution is the linear operation, the output sequence y(n) is simply
the summation of all segments expressed as
y(n) =
¸
m
y
m
(n): (7:3:23)
Because each output segment y
m
(n) overlaps the following segment y
m÷1
(n) by L
samples, (7.3.23) implies the actual addition of the last L samples in segment y
m
(n)
with the first L samples in segment y
m÷1
(n).
This efficient FIR filtering using the overlap-add technique is implemented by the
MATLAB function
y = fftfilt(h, x);
or
y = fftfilt(h, x, N);
The fftfilt function filters the input signal in the vector x with the FIR filter
described by the coefficient vector h. The function y = fftfilt(h, x)chooses an
FFT and a data block length that automatically guarantees efficient execution time.
However, we can specify the FFT length N by using y = fftfilt(h, x, N).
7.3.5 Spectrogram
The PDS introduced in Section 7.3.3 is a powerful technique to show how the power of
the signal is distributed among the various frequency components. However, this
method will result in a distorted (blurred) spectrum when the signal is non-stationary.
332 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
For a time-varying signal, it is more useful to compute a local spectrum that measures
spectral contents over a short time interval.
In this section, we use a sliding window defined in (7.3.17) to break up a long
sequence into several short finite-length blocks of N samples x
/
m
(n), and then perform
the FFT to obtain the time-dependent frequency spectrum at each short segment to
obtain
X
m
(k) =
¸
N÷1
n=0
x
/
m
(n)W
kn
N
, k = 0, 1, . . . , N ÷1: (7:3:24)
This process is repeated for the next block of N samples as illustrated in Figure 7.12.
This technique is also called the short-term Fourier transform, since X
m
(k) is just the
DFT spectrum of the short segment of x
m
(n) that lies inside the sliding window w(n).
This form of time-dependent Fourier transform has several applications in speech,
sonar, and radar signal processing.
Equation (7.3.24) shows that X
m
(k) is a two-dimensional sequence. The index k
represents frequency as defined in (7.3.1), and the block index m represents time.
Since the result is a function of both time and frequency, a three-dimensional graphical
display is needed. This is done by plotting [X
m
(k)[ as a function of both k and m using
gray-scale (or color) images. The resulting three-dimensional graphic is called the
spectrogram. It uses the x-axis to represent time and the y-axis to represent frequency.
The gray level (or color) at point (m, k) is proportional to [X
m
(k)[. The large values are
black, and the small ones are white.
The Signal Processing Toolbox provides a function, specgram, to compute spectro-
gram. This MATLAB function has the form
B = specgram(a, nfft, Fs, window, noverlap);
where B is a matrix containing the complex spectrogram values [X
m
(k)[, and other
arguments are defined in the function psd. It is common to pick the overlap to
be around 50 percent as shown in Figure 7.12. The specgram function with no
output arguments displays the scaled logarithm of the spectrogram in the current
graphic window. See the Signal Processing Toolbox for Use with MATLAB [7]
for details.
7.4 Implementation Considerations
The FFT algorithm can be realized as a program in a general-purpose computer, a DSP
processor, or implemented in special-purpose hardware. Many FFT routines are avail-
able in C and assembly programs. We probably would not even have to write an FFT
routine for a given application. However, it is important to understand the implementa-
tion issues in order to use FFT properly. In this section, we only consider the radix-2
FFT algorithms.
IMPLEMENTATION CONSIDERATIONS 333
7.4.1 Computational Issues
As illustrated in Figure 7.5, the radix-2 FFT algorithm takes two input samples at a time
from memory, performs the butterfly computations, and returns the resulting numbers
to the same input memory locations. This process is repeated N log
2
N times in the
computation of an N-point FFT. The FFT routines accept complex-valued inputs,
therefore the number of memory locations required is 2N. Complex-valued signals are
quite common in communications such as modems. However, most signals such as
speech are real-valued. To use the available FFT routine, we have to set the imaginary
part of each sample value to 0 for real input data. Note that each complex multiplication
is of the form
(a ÷jb)(c ÷jd) = (ac ÷bd) ÷j(bc ÷ad)
and therefore requires four real multiplications and two real additions.
The number of multiplications and the storage requirements can be reduced if the
signal has special properties. For example, if x(n) is real, only N/2 samples from X(0) to
X(N/2) need to be computed as shown by complex-conjugated property (7.1.15). In
addition, if x(n) is an even function of n, only the real part of X(k) is non-zero. If x(n) is
odd, only the imaginary part is non-zero.
The computation of twiddle factors W
kn
N
usually takes longer than the computation of
complex multiplications. In most FFT programs on general-purpose computers, the
sine and cosine calculations defined in (7.1.4) are embedded in the program for con-
venience. If N is fixed, it is preferable to tabulate the values of twiddle factors so that
they can be looked up during the computation of FFT algorithms. When the FFT is
performed repeatedly with N being constant, the computation of twiddle factors need
not be repeated. In addition, in an efficient implementation of FFT algorithm on a DSP
processor, the twiddle factors are computed once and then stored in a table during the
programming stage.
There are other implementation issues such as indexing, bit reversal, and parallelism
in computations. The complexity of FFT algorithms is usually measured by the required
number of arithmetic operations (multiplications and additions). However, in practical
implementations on DSP chips, the architecture, instruction set, data structures, and
memory organizations of the processors are critical factors. Modern DSP chips such as
the TMS320C55x usually provide single-cycle multiplication-and-accumulation oper-
ation, bit-reversal addressing, and a high degree of instruction parallelism to efficiently
implement FFT algorithms. These issues will be discussed further in Section 7.5.
7.4.2 Finite-Precision Effects
Since FFT is often employed in DSP hardware for real-time applications, it is important
to analyze the finite-precision effects in FFT computations. We assume that the FFT
computations are being carried out using fixed-point arithmetic. With clever scaling and
checking for overflow, the most critical error in the computation is due to roundoff
errors. Without loss of generality, we analyze the decimation-in-time radix-2 FFT
algorithm introduced in Section 7.2.1.
334 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
From the flow-graph of the FFT algorithm shown in Figure 7.6, X(k) are computed
by a series of butterfly computations with a single complex multiplication per butterfly
network. Note that some of the butterfly computations require multiplications by
÷
÷
1
(such as 2-point FFT in the first stage) that do not require multiplication in practical
implementation, thus avoiding roundoff errors.
Figure 7.6 shows that the computation of N-point FFT requires M = log
2
N stages.
There are N/2 butterflies in the first stage, N/4 in the second stage, and so on. Thus the
total number of butterflies required to produce an output sample is
N
2
÷
N
4
÷ ÷2 ÷1 = 2
M÷1
÷2
M÷2
÷ ÷2 ÷1
= 2
M÷1
1 ÷
1
2

÷ ÷
1
2

M÷1
¸ ¸
= 2
M÷1
¸
M÷1
m=0
1
2

m
= 2
M
1 ÷
1
2

M
¸ ¸
= N ÷1:
(7:4:1)
The quantization errors introduced at the mth stage appear at the output after propaga-
tion through (m÷1) stages, while getting multiplied by the twiddle factors at each
subsequent stage. Since the magnitude of the twiddle factor is always unity, the vari-
ances of the quantization errors do not change while propagating to the output. If we
assume that the quantization errors in each butterfly are uncorrelated with the errors in
other butterflies, the total number of roundoff error sources contributing to the output
is 4(N ÷1). Therefore the variance of the output roundoff error is
s
2
e
= 4(N ÷1)
2
÷2B
12
~
N2
÷2B
3
: (7:4:2)
As mentioned earlier, some of the butterflies do not require multiplications in practical
implementation, thus the total roundoff error is less than the one given in (7.4.2).
The definition of DFT given in (7.1.3) shows that we can scale the input sequence
with the condition
[x(n)[ <
1
N
(7:4:3)
to prevent the overflow at the output because [e
÷j(2p=N)kn
[ = 1. For example, in a 1024-
point FFT, the input data must be shifted right by 10 bits. If the original data is 16-bit,
the effective wordlength after scaling is reduced to only 6 bits. This worst-case scaling
substantially reduces the resolution of the FFT results.
Instead of scaling the input samples by 1/N at the beginning, we can scale the signals
at each stage since the FFT algorithm consists of a sequence of stages. Figure 7.5 shows
that we can avoid overflow within the FFT by scaling the input at each stage by 1/2
(right shift one bit in a fixed-point hardware) because the outputs of each butterfly
involve the addition of two numbers. That is, we shift right the input by 1 bit, perform
the first stage of FFT, shift right that result by 1 bit, performthe second stage of FFT, and
so on. This unconditional scaling process does not affect the signal level at the output of
IMPLEMENTATION CONSIDERATIONS 335
the FFT, but it significantly reduces the variance of the quantization errors at the output.
Thus it provides a better accuracy than unconditional scaling the input by 1/N.
An alternative conditional scaling method examines the results of each FFT stage to
determine whether all the results of that stage should be scaled. If all the results in a
particular stage have magnitude less than 1, no scaling is necessary at that stage.
Otherwise, all the inputs of that stage have to be scaled by 1/2. The conditional scaling
technique achieves much better accuracy since we may scale less often than the uncon-
ditional scaling method. However, this conditional scaling method increases software
complexity and may require longer execution time.
7.5 Experiments Using the TMS320C55x
In this section, we implement the decimation-in-time FFT algorithm using the floating-
point C, fixed-point C, and assembly language. We then implement the IFFT algorithm
using the same FFT routine. Finally, we apply both the FFT and IFFT for fast
convolution.
7.5.1 Experiment 7A ± Radix-2 Complex FFT
The decimation-in-time FFT algorithm based on (7.2.6) shows how to compute an N-
point DFT by combining N/2-point DFT sections. The computation described by
(7.2.6) is called the butterfly computation and is shown graphically in Figure 7.5. The
floating-point C function for computing a radix-2 complex decimation-in-time FFT is
listed in Table 7.2. This program will compute an N-point FFT using a sinusoidal signal
as input. The output of this program should be all zeros except at the FFT bins of X(k)
and X(N ÷k). By changing the constants N and EXP, we are able to perform different
length of radix-2 complex FFT using routine shown in Table 7.3. Since this is a complex
radix-2 FFT routine, the imaginary portion of the complex data buffer must be set to 0
if the data is real.
Table 7.2 Floating-point C program for testing the FFT algorithm
/*
Example to test floating-point complex FFT
*/
#include <math.h>
#include "fcomplex.h" /* Floating-point complex.h header file */
#include "input7_f.dat" /* Floating-point testing data */
extern void fft(complex *, unsigned int, complex *, unsigned int);
extern void bit_rev(complex *X, int M);
#define N 128 /* FFT size */
#define EXP 7 /* EXP = log2(N)*/
#define pi 3.1415926535897
336 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
Table 7.2 (continued )
complex X[N]; /* Declare input buffer */
complex W[EXP]; /* Twiddle factors table */
complex temp;
float spectrum[N];
float re1[N],im1[N];
void main()
{
unsigned int i, j, L, LE, LE1;
for(L = 1; L <= EXP; L÷÷) /* Create twiddle-factor table */
{
LE = 1<L; /* LE = 2^L = points of sub DFT */
LE1 = LE1; /* Number of butterflies in sub DFT */
W[L÷1].re = cos(pi/LE1);
W[L÷1].im = ÷sin(pi/LE1);
}
j = 0;
for(;;)
{
for(i = 0; i < N; i÷÷)
{
/* Generate input samples */
X[i].re = input7_f[j++];
X[i].im = 0.0;
/* Copy to reference buffer */
re1[i]= X[i].re;
im1[i]= X[i].im;
if(j == 1664)
j = 0;
}
/* Start FFT */
bit_rev(X, EXP); /* Arrange X[]in bit-reversal order */
fft(X, EXP, W, 1); /* Perform FFT */
/* Verify FFT results */
for(i = 0; i < N; i÷÷)
{
/* Compute power spectrum */
temp.re = X[i].re*X[i].re;
temp.im = X[i].im*X[i].im;
spectrum[i]=(temp.re ÷ temp.im)*4;
}
}
}
EXPERIMENTS USING THE TMS320C55X 337
The complex radix-2 FFT program listed in Table 7.3 computes the complex decima-
tion-in-time FFT algorithm as shown in Figure 7.5. To prevent the results from over-
flowing, the intermediate results are scaled down in each stage as described in Section
7.4.2. The radix-2 FFT function contains two complex arguments and two unsigned
integer arguments. They are the complex input sample, X [N], the power of the radix-2
FFT, EXP, the initial complex twiddle-factor table W [EXP], and the scaling flag SCALE.
The FFT is performed in place, that is, the complex input array is overwritten by
the output array. The initial twiddle factors are created by the C function listed in
Table 7.2.
As discussed in Section 7.2.1, the data used for FFTneed to be placed in the bit-reversal
order. An N-point FFTbit-reversal example is given in Table 7.1. Table 7.4 illustrated the
C function that performs the bit-reversal addressing task.
Table 7.3 Floating-point complex radix-2 FFT routine in C
/*
fft_float.c ÷ Floating-point complex radix-2 decimation-in-time FFT
Perform in-place FFT, the output overwrite the input buffer
*/
#include "fcomplex.h" /* Floating-point complex.h header file */
void fft(complex *X, unsigned int M, complex *W, unsigned int SCALE)
{
complex temp; /* Temporary storage of complex variable */
complex U; /* Twiddle factor W^k */
unsigned int i,j;
unsigned int id; /* Index for lower point in butterfly */
unsigned int N = 1<EXP; /* Number of points for FFT */
unsigned int L; /* FFT stage */
unsigned int LE; /* Number of points in sub FFT at stage
L and offset to next FFT in stage */
unsigned int LE1; /* Number of butterflies in one FFT at
stage L. Also is offset to lower
point in butterfly at stage L */
float scale;
scale = 0.5;
if (SCALE == 0)
scale = 1.0;
for(L = 1; L <= EXP; L÷÷) /* FFT butterfly */
{
LE = 1<L; /* LE = 2^L = points of sub DFT */
LE1 = LE1; /*Number of butterflies in sub-DFT */
U.re = 1.0;
U.im = 0.;
338 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
Table 7.3 (continued )
for(j = 0; j < LE1; j÷÷)
{
for(i = j; i < N; i ÷= LE) /* Do the butterflies */
{
id = i÷LE1;
temp.re =(X[id].re*U.re ÷ X[id].im*U.im)*scale;
temp.im =(X[id].im*U.re ÷ X[id].re*U.im)*scale;
X[id].re = X[i].re*scale ÷ temp.re;
X[id].im = X[i].im*scale ÷ temp.im;
X[i].re = X[i].re*scale ÷ temp.re;
X[i].im = X[i].im*scale ÷ temp.im;
}
/* Recursive compute W^k as U*W^(k÷1) */
temp.re = U.re*W[L÷1].re ÷ U.im*W[L÷1].im;
U.im = U.re*W[L÷1].im ÷ U.im*W[L÷1].re;
U.re = temp.re;
}
}
}
Table 7.4 Floating-point bit-reversal function
/*
fbit_rev.c ÷ Arrange input samples in bit-reversal order
The index j is the bit-reversal of i
*/
#include "fcomplex.h" /* Floating-point complex.h header file */
void bit_rev(complex *X, unsigned int EXP)
{
unsigned int i, j, k;
unsigned int N = 1<EXP; /* Number of points for FFT */
unsigned int N2 = N1;
complex temp; /* Temp storage of the complex variable */
for(j = 0, i = 1; i < N÷1; i÷÷)
{
k = N2;
while(k <= j)
{
continues overleaf
EXPERIMENTS USING THE TMS320C55X 339
Table 7.4 (continued )
j ÷= k;
k = 1;
}
j ÷= k;
if(i < j)
{
temp = X[j];
X[j]= X[i];
X[i]= temp;
}
}
}
Based on the floating-point C function, the fixed-point conversion is done by employ-
ing the C intrinsic functions. The implementation of the decimation-in-time FFT butter-
fly computation using the C55x intrinsics is listed below:
ltemp.re = _lsmpy(X[id].re, U.re);
temp.re = (_smas(ltemp.re, X[id].im, U.im) 16);
temp.re = _sadd(temp.re, 1) scale; /* Rounding & scale */
ltemp.im = _lsmpy(X[id].im, U.re);
temp.im = (_smac(ltemp.im, X[id].re, U.im) 16);
temp.im = _sadd(temp.im, 1) scale; /* Rounding & scale */
X[id].re = _ssub(X[i].re scale, temp.re);
X[id].im = _ssub(X[i].im scale, temp.im);
X[i].re = _sadd(X[i].re scale, temp.re);
X[i].im = _sadd(X[i].im scale, temp.im);
In the program, X[]is the complex sample buffer and U is the complex twiddle factor.
The scale is done by right-shifting 1-bit instead of multiply by 0.5.
Go through the following steps for Experiment 7A:
1. Verify the floating-point C programs test_fft.c, fft_float.c, fbit_rev.c,
and fcomplex.h using a PC or C55x simulator. The output should be all zeros
except X(k) and X(N÷k). These squared values are equal to 1.0. The floating-point
program will be used as reference for the code development using the fixed-point
C and assembly language. The floating-point program uses the floating-point data
file input7_f.dat, while the fixed-point data file input7_i.dat will be used for
the rest of experiments.
2. Create the project exp7a using CCS. Add the command file exp7.cmd, the
functions epx7a.c, fft_a.c, and ibit_rev.c, and the header file
icomplex.h from the software package into the project.
340 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
3. Build the fixed-point FFT project and verify the results. Comparing the results
with the floating-point complex radix-2 FFT results obtained by running it on
the PC.
4. The FFT output samples are squared and placed in a data buffer named
spectrum []. Use CCS to plot the results by displaying the spectrum []and
re1 []buffer.
5. Profile the DSP run-time clock cycles for 128-point and 1024-point FFTs. Record
the memory usage of the fixed-point functions bit_rev()and fft().
7.5.2 Experiment 7B ± Radix-2 Complex FFT Using Assembly
Language
Although using intrinsics can improve the DSP performance, the assembly language
implementation has been proven to have the fastest execution speed and memory
efficiency for most applications, especially for computational intensive algorithms
such as FFT. The development time for assembly code, however, will be much
longer than that of C code. In addition, the maintenance and upgrade of assembly
code are usually more difficult. In this experiment, we will use C55x assembly routines
for computing the same radix-2 FFT algorithm as the fixed-point C function used
in Experiment 7A. The assembly FFT routine is listed in Table 7.5. This routine is
written based on the C function used for Experiment 7A, and it follows the C55x
C calling convention. For readability, the assembly code has been written to mimic
the C function of Experiment 7A closely. It optimizes the memory usage but not
the run-time efficiency. By unrolling the loop and taking advantage of the FFT butterfly
characteristics, the FFT execution speed can be further improved with the expense of
the memory space, see the exercise problems at the end of this chapter.
In fft.asm, the local variables are defined as structure using the stack relative
addressing mode when the assembly routine is called. The last memory location con-
tains the return address of the caller function. Since the status registers ST1 and ST3 will
be modified, we use two stack locations to store the contents of these status registers at
entry. The status registers will be restored upon returning to the caller function. The
complex temporary variable is stored in two consecutive memory locations by using a
bracket with the numerical number to indicate the number of memory locations for the
integer data type.
The FFT implementation is carried out in three nested loops. The butterfly computa-
tion is implemented in the inner loop and the group loop is in the middle, while the
stages are managed by the outer loop. Among these three loops, the butterfly loop is
repeated most often. We use the local block repeat instruction, rptblocal, for the
butterfly loop and the middle loop to minimize the loop overhead. We also use parallel
instructions, modulo addressing, and dual memory access instructions to further
improve the efficiency of butterfly computation. By limiting the size of the loop, we
can place the middle loop inside the DSP instruction buffer queue (IBQ) as well. The
FFT computation is improved since the two inner loops are only fetched once from the
program memory each time we compute the groups and butterflies. The twiddle-factor
EXPERIMENTS USING THE TMS320C55X 341
Table 7.5 FFT routine using the C55x assembly language
outer_loop ; for(L = 1; L <= M; L÷÷)
mov fft.d_L,T0 ; Note: Since the buffer is
[ [ mov #2, AC0 ; arranged in re,im pairs
sfts AC0, T0 ; the index to the buffer
neg T0 ; is doubled
[ [ mov fft.d_N, AC1 ; But the repeat counters
sftl AC1,T0 ; are not doubled
mov AC0,T0 ; LE = 2< L
[ [ sfts AC0,#÷1
mov AC0,AR0 ; LE1 = LE1
[ [ sfts AC0,#÷1
sub #1,AC0 ; Initialize mid_loop counter
mov mmap(AC0L),BRC0 ; BRC0 = LE1÷1
sub #1,AC1 ; Initialize inner loop counter
mov mmap(AC1L),BRC1 ; BRC1 =(NL)÷1
add AR1,AR0
mov #0,T2 ; j = 0
[ [ rptblocal mid_loop-1 ; for(j = 0; j < LE1; j÷÷)
mov T2,AR5 ; AR5 = id = i÷LE1
mov T2,AR3
add AR0,AR5 ; AR5 = pointer to X[id].re
add #1,AR5,AR2 ; AR2 = pointer to X[id].im
add AR1,AR3 ; AR3 = pointer to X[i].re
[ [ rptblocal inner_loop-1 ; for(i = j; i < N; i ÷= LE)
mpy *AR5÷, *CDP÷, AC0 ; AC0 = (X[id].re*U.re
:: mpy *AR2÷, *CDP+, AC1 ; ÷X[id].im*U.im)/SCALE
masr *AR5÷, *CDP÷, AC0 ; AC1 = (X[id].im*U.re
:: macr *AR2÷, *CDP÷, AC1 ; ÷X[id].re*U.im)/SCALE
mov pair(hi(AC0)),dbl(*AR4); AC0H = temp.re AC1H = temp.im
[ [ mov dbl(*AR3), AC0
xcc scale,TC1
[ [ mov AC0 #1,dual(*AR3) ; Scale X[i]by 1/SCALE
mov dbl(*AR3), AC0
scale
add T0,AR2
[ [ sub dual(*AR4),AC0,AC1 ; X[id].re = X[i].re/SCALE-temp.re
mov AC1,dbl(*(AR5÷T0)) ; X[id].im = X[i].im/SCALE-temp.im
[ [ add dual(*AR4),AC0 ; X[i].re = X[i].re/SCALE+temp.re
mov AC0,dbl(*(AR3÷T0)) ; X[i].im = X[i].im/SCALE÷temp.im
inner_loop ; End of inner loop
amar *CDP÷
amar *CDP÷ ; Update k for pointer to U[k]
[ [ add #2,T2 ; Update j
mid_loop ; End of mid-loop
sub #1,T1
add #1,fft.d_L ; Update L
bcc outer_loop,T1 > 0 ; End of outer-loop
342 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
table is pre-calculated during the initialization phase. The calculation of the twiddle
factors can be implemented as follows:
for(i = 0, l = 1; l <= EXP; l÷÷)
{
SL = 1<l; /* LE = 2^L = points of sub FFT */
SL1 = SL1; /* # of twiddle factors in sub-FFT */
for(j = 0; j < SL1; j÷÷)
{
W.re = (int)((0x7fff*cos(j*pi/SL1)) ÷ 0.5);
W.im = ÷(int)((0x7fff*sin(j*pi/SL1))÷ 0.5);
U [i÷÷]= W;
}
}
where U[]and W are defined as complex data type.
The bit-reversal addressing assembly routine is shown in Table 7.6. Since we use the
special C55x bit-reversal addressing mode, the assembly routine is no longer the same as
the C function used in the previous experiment. The bit-reversal addressing mode uses
the syntax of *(AR1÷T0B), where the temporary register T0 contains the size of the
buffer, and the letter B indicates that the memory addressing is in a bit-reversal order.
For Experiment 7B, we also use the data and program pragma directives to manage the
memory space.
Complete the following steps for Experiment 7B:
1. Create the project exp7b, add files exp7.cmd, exp7b.c, w_table.c, fft.asm,
and bit_rev.asm from the software package into the project.
2. Build and verify the FFT function, and compare the results with the results obtained
from Experiment 7A. Make sure that the scale flag for the FFT routine is set to 1.
3. Profile the FFT run-time clock cycles and its memory usage again and compare
these results with those obtained in Experiment 7A.
Table 7.6 Bit-reversal routine in the C55x assembly language
rptblocal loop_end-1 ; Start bit-reversal loop
mov dbl(*AR0),AC0 ; Get a pair of samples
[ [ amov AR1,T1
mov dbl(*AR1),AC1 ; Get another pair
[ [ asub AR0,T1
xccpart swap1,T1 >= #0
[ [ mov AC1,dbl(*AR0+) ; Swap samples if j >= i
swap1
xccpart loop_end,T1 >= #0
[ [ mov AC0,dbl(*(AR1÷T0B))
loop_end ; End bit-reversal loop
EXPERIMENTS USING THE TMS320C55X 343
7.5.3 Experiment 7C ± FFT and IFFT
As discussed in Section 7.2.3, the inverse DFT defined by (7.2.11) is similar to the DFT
defined in (7.1.6). Thus the FFT routine developed in Experiment 7B can be modified
for computing the inverse FFT. Two simple changes are needed in order to use the same
FFT routine for the IFFT calculation. First, the conjugating twiddle factors imply the
sign change of the imaginary portion of the complex samples. That is, X[i].im =
÷X[I].im. Second, the normalization of 1=N is handled in the FFT routine by
setting the scale flag to 0. Table 7.7 shows the example of computing both the FFT
and IFFT.
Go over the following steps for Experiment 7C:
1. Create the project epx7c and include the files exp7.cmd, exp7c.c, w_table.c,
fft.asm, and bit_rev.asm from the software package into the project.
2. Build and view the IFFT results by plotting and comparing the input array re1 []
and IFFT output array re2 []. Make sure that the scale flag for the FFT calcula-
tion is set to 1 (one), and the IFFT calculation is set to 0 (zero).
7.5.4 Experiment 7D ± Fast Convolution
As discussed in Section 7.3.4, the application of fast convolution using FFT/IFFT is the
most efficient technique of FIR filtering for long time-domain sequence such as for
high-fidelity digital audio systems, or FIR filtering in frequency-domain such as in the
xDSL modems. The fast convolution algorithm is shown in Figure 7.13. There are two
basic methods for FFT convolution as mentioned in Section 7.3.4. This experiment will
use the overlap-add technique. This method involves the following steps:
± Pad M = N ÷L zeros to the FIR filter impulse response of length L where N > L,
and process the sequence using an N-point FFT. Store the results in the complex
buffer H[N].
Table 7.7 Perform FFT and IFFT using the same routine
/* Start FFT */
bit_rev(X,EXP); /* Arrange X []in bit-reversal order */
fft(X,EXP,U,1); /* Perform FFT */
/* Inverse FFT */
for(i = 0; i < N; i÷÷) /* Change the sign of imaginary part */
{
X [i].im = ÷X [i].im;
}
bit_rev(X,EXP); /* Arrange sample in bit-reversal order */
fft(X,EXP,U,0); /* Perform IFFT */
344 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
± Segment the input sequence of length M with L ÷1 zeros padded at the end.
± Process each segment of data samples with an N-point FFT to obtain the complex
array X [N].
± Multiply H and X in frequency domain to obtain Y.
± Perform N-point IFFT to find the time-domain filtered sequence.
± Add the first L samples that are overlapped with the previous segment to form the
output. All resulting segments are combined to obtain y(n).
The C program implementation of fast convolution using FFT and IFFT is listed in
Table 7.8, where we use the same data file and FIR coefficients as the experiments given
in Chapter 5. In general, for low- to median-order FIR filters, the direct FIR routines
introduced in Chapter 5 are more efficient. Experiment 5A shows that an FIR filter can
be implemented as one clock cycle per filter tap, while Experiments 5B and 5C complete
two taps per cycle. However, the computational complexity of those routines is linearly
increased with the number of coefficients. When the application requires high-order
FIR filters, the computation requirements can be reduced by using fast convolution as
shown in this experiment.
Table 7.8 Fast convolution using FFT and IFFT
/* Initialization */
for(i = 0; i < L÷1; i÷÷) /*Initialize overlap buffer */
OVRLAP [i]= 0;
for(i = 0; i < L; i÷÷) /* Copy filter coefficients to buffer */
{
X [i].re = LP_h [i];
X [i].im = 0;
}
for(i = i; i < N; i÷÷) /* Pad zeros to the buffer */
{
X [i].re = 0;
X [i].im = 0;
}
w_table(U,EXP); /* Create twiddle-factor table */
bit_rev(X,EXP); /* Bit-reversal arrangement of coefficients */
fft(X,EXP,U,1); /* FFT of filter coefficients */
for(i = 0; i < N; i÷÷) /* Save frequency-domain coefficients */
{
H [i].re = X [i].re < EXP;
H [i].im = X [i].im < EXP;
}
continues overleaf
EXPERIMENTS USING THE TMS320C55X 345
Table 7.8 (continued )
/* Start fast convolution test */
j = 0;
for(;;)
{
for(i = 0; i < M; i÷÷)
{
X [i].re = input [j÷÷]; / * Generate input samples */
X [i].im = 0;
if(j==160)
j = 0;
}
for(i = i; i < N; i÷÷) /* Fill zeros to data buffer */
{
X [i].re = 0;
X [i].im = 0;
}
/* Start FFT convolution */
bit_rev(X, EXP); /* Samples in bit-reversal order */
fft(X, EXP, U, 1); /* Perform FFT */
freqflt(X, H, N); /* Frequency domain filtering */
bit_rev(X, EXP); /* Samples in bit-reversal order */
fft(X, EXP, U, 0); /* Perform IFFT */
olap_add(X, OVRLAP, L, M, N); /* Overlap-add algorithm */
}
Go through the following steps for Experiment 7D:
1. Create the project exp7d, add the files exp7.cmd, exp7d.c, w_table.c,
fft.asm, bit_rev.asm, freqflt.asm, and olap_add.asm from the software
package to the project.
2. Build and verify the fast convolution results, and compare the results with the
results obtained in Experiment 5A.
3. Profile the run-time clock cycles of the fast convolution using FFT/IFFT
for various FIR filter lengths by using different filter coefficient files
firlp8.dat, firlp16.dat, firlp32.dat, firlp64.dat, firlp128.dat,
firlp256.dat, and firlp512.dat. These files are included in the experiment
software package.
References
[1] D. J. DeFatta, J. G. Lucas, and W. S. Hodgkiss, Digital Signal Processing: A System Design
Approach, New York: Wiley, 1988.
346 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
[2] N. Ahmed and T. Natarajan, Discrete-Time Signals and Systems, Englewood Cliffs, NJ: Prentice-
Hall, 1983.
[3] V. K. Ingle and J. G. Proakis, Digital Signal Processing Using MATLAB V.4, Boston: PWS
Publishing, 1997.
[4] L. B. Jackson, Digital Filters and Signal Processing, 2nd Ed., Boston: Kluwer Academic, 1989.
[5] MATLAB User's Guide, Math Works, 1992.
[6] MATLAB Reference Guide, Math Works, 1992.
[7] Signal Processing Toolbox for Use with MATLAB, Math Works, 1994.
[8] A. V. Oppenheim and R. W. Schafer, Discrete-Time Signal Processing, Englewood Cliffs, NJ:
Prentice-Hall, 1989.
[9] S. J. Orfanidis, Introduction to Signal Processing, Englewood Cliffs, NJ: Prentice-Hall, 1996.
[10] J. G. Proakis and D. G. Manolakis, Digital Signal Processing ± Principles, Algorithms, and
Applications, 3rd Ed., Englewood Cliffs, NJ: Prentice-Hall, 1996.
[11] A Bateman and W. Yates, Digital Signal Processing Design, New York: Computer Science Press,
1989.
[12] S. D. Stearns and D. R. Hush, Digital Signal Analysis, 2nd Ed., Englewood Cliffs, NJ: Prentice-
Hall, 1990.
Exercises
Part A
1. Compute the four-point DFT of the sequence ¦1, 1, 1, 1¦ using the matrix equations given in
(7.1.7) and (7.1.8).
2. Repeat Problem 1 with eight-point DFT of sequence ¦1, 1, 1, 1, 0, 0, 0, 0¦. Compare the results
with the results of Problem 1.
3. Calculate the DFTs of the following signals:
(a) x(n) = d(n).
(b) x(n) = d(n ÷n
0
), 0 < n
0
< N.
(c) x(n) = c
n
, 0 _ n _ N ÷1.
(d) x(n) = cos(!
0
n), 0 _ n _ N.
(e) x(n) = sin(!
0
n), 0 _ n _ N.
4. Prove the symmetry and periodicity properties of the twiddle factors defined as
(a) W
k÷N=2
N
= ÷W
k
N
.
(b) W
k÷N
N
= W
k
N
.
5. Generalize the derivation of Example 7.7 to a four-point DFT and show a detailed signal-flow
graph of four-point DFT.
6. Consider the following two sequences:
x
1
(n) = x
2
(n) = 1, 0 _ n _ N ÷1:
EXERCISES 347
(a) Compute the circular convolution of the two sequences using DFT and IDFT.
(b) Show that the linear convolution of these two sequences is the triangular sequence given
by
x
3
(n) =
n ÷1, 0 _ n < N
2N ÷n ÷1, N _ n < 2N
0, otherwise.

(c) How to make the circular convolution of the two sequences becomes a triangular
sequence defined in (b)?
7. Construct the signal-flow diagram of FFT for N = 16 using the decimation-in-time method
with bit-reversal input.
8. Construct the signal-flow diagram of FFT for N = 8 using the decimation-in-frequency
method without bit-reversal input.
9. Complete the development of decimation-in-frequency FFT algorithm for N = 8. Show the
detailed flow graph.
10. Consider a digitized signal of one second with the sampling rate 20 kHz. The spectrum is
desired with a computational frequency resolution of 100 Hz or less. Is this possible? If
possible, what FFT size N should be used?
11. A 1 kHz sinusoid is sampled at 8 kHz. The 128-point FFT is performed to compute X(k). At
what frequency indices k we expect to observe any peaks in [X(k)[?
12. A touch-tone phone with a dual-tone multi-frequency (DTMF) transmitter encodes each
keypress as a sum of two sinusoids, with one frequency taken from each of the following
groups:
Vertical group: 697, 770, 852, 941 Hz
Horizontal group: 1209, 1336, 1477, 1633 Hz
What is the smallest DFT size N that we can distinguish these two sinusoids from the
computed spectrum? The sampling rate used in telecommunications is 8 kHz.
13. Compute the linear convolution y(n) = x(n)
+
h(n) using 512-point FFT, where x(n) is of
length 4096 and h(n) is of length 256.
(a) How many FFTs and how many adds are required using the overlap-add method?
(b) How many FFTs are required using the overlap-save method?
(c) What is the length of output y(n)?
348 FAST FOURIER TRANSFORM AND ITS APPLICATIONS
Part B
14. Write a C or MATLAB program to compute the fast convolution of a long sequence with a
short sequence employing the overlap-save method introduced in Section 7.3.4. Compare the
results with the MATLAB function fftfilt that use overlap-add method.
15. Experiment with the capability of the psd function in the MATLAB. Use a sinusoid
embedded in white noise for testing signal.
16. Using the MATLAB function specgram to display the spectrogram of the speech file
timit1.asc included in the software package.
Part C
17. The radix-2 FFT code used in the experiments is written in consideration of minimizing
the code size. An alternative FFT implementation can be more efficient in terms of the
execution speed with the expense of using more program memory locations. For example,
the twiddle factors used by the first stage and the first group of other stages are
constants, W
0
N
= 1. Therefore the multiplication operations in these stages can be simplified.
Modify the assembly FFT routine given in Table 7.5 to incorporate this observation. Profile
the run-time clock cycles and record the memory usage. Compare the results with those
obtained by Experiment 7C.
18. The radix-2 FFT is the most widely used algorithm for FFT computation. When the number
of data samples are a power of 2n (i.e., N = 2
2n
= 4
n
), we can further improve the run-time
efficiency by employing the radix-4 FFT algorithm. Modify the assembly FFT routine give
in Table 7.5 for the radix-4 FFT algorithm. Profile the run-time clock cycles, and record the
memory space usage for a 1024-point radix-4 FFT (2
10
= 4
5
= 1024). Compare the radix-4
FFT results with the results of 1024-point radix-2 FFT computed by the assembly routine.
19. Take advantage of twiddle factor, W
0
N
= 1, to further improve the radix-4 FFT algorithm
run-time efficiency. Compare the results of 1024-point FFT implementation using different
approaches.
20. Most of DSP applications have real input samples, our complex FFT implementation zeros
out the imaginary components of the complex buffer (see exp7c.c). This approach is simple
and easy, but it is not efficient in terms of the execution speed. For real input, we can split the
even and odd samples into two sequences, and compute both even and odd sequences in
parallel. This approach will reduce the execution time by approximately 50 percent. Given a
real value input x(n) of 2N samples, we can define c(n) = a(n) ÷jb(n), where two inputs
a(n) = x(n) and b(n) = x(n ÷1) are real sequences. We can represent these sequences
as a(n) = [c(n) ÷c
+
(n)[=2 and b(n) = ÷j[c(n) ÷c
+
(n)[=2, then they can be written in terms
of DFTs as and A(k) = [C(k) ÷C
+
(N ÷k)[=2 and B(k) = ÷j[C(k) ÷C
+
(N ÷k)[=2.
Finally, the real input FFT can be obtained by X(k) = A(k) ÷W
k
2N
B(k) and
X(k ÷N) = A(k) ÷W
k
2N
B(k), where k = 0, 1, . . . , N ÷1. Modify the complex radix-2 FFT
assembly routine to efficiently compute 2N real input samples.
EXERCISES 349
8
Adaptive Filtering
As discussed in previous chapters, filtering refers to the linear process designed to alter
the spectral content of an input signal in a specified manner. In Chapters 5 and 6, we
introduced techniques for designing and implementing FIR and IIR filters for given
specifications. Conventional FIR and IIR filters are time-invariant. They perform linear
operations on an input signal to generate an output signal based on the fixed coeffi-
cients. Adaptive filters are time varying, filter characteristics such as bandwidth and
frequency response change with time. Thus the filter coefficients cannot be determined
when the filter is implemented. The coefficients of the adaptive filter are adjusted
automatically by an adaptive algorithm based on incoming signals. This has the import-
ant effect of enabling adaptive filters to be applied in areas where the exact filtering
operation required is unknown or is non-stationary.
In Section 8.1, we will review the concepts of random processes that are useful in the
development and analysis of various adaptive algorithms. The most popular least-mean-
square (LMS) algorithmwill be introduced in Section 8.2. Its important properties will be
analyzed in Section 8.3. Two widely used modified adaptive algorithms, the normalized
and leaky LMS algorithms, will be introduced in Section 8.4. In this chapter, we introduce
and analyze the LMS algorithm following the derivation and analysis given in [8]. In
Section 8.5, we will briefly introduce some important applications of adaptive filtering.
The implementation considerations will be discussed in Section 8.6, and the DSP imple-
mentations using the TMS320C55x will be presented in Section 8.7.
8.1 Introduction to Random Processes
A signal is called a deterministic signal if it can be described precisely and be reproduced
exactly and repeatedly. However, the signals encountered in practice are not necessarily
of this type. A signal that is generated in a random fashion and cannot be described by
mathematical expressions or rules is called a random (or stochastic) signal. The signals
in the real world are often random in nature. Some common examples of random
signals are speech, music, and noises. These signals cannot be reproduced and need to
be modeled and analyzed using statistical techniques. We have briefly introduced
probability and random variables in Section 3.3. In this section, we will review the
important properties of the random processes and introduce fundamental techniques
for processing and analyzing them.
Real-Time Digital Signal Processing. Sen M Kuo, Bob H Lee
Copyright #2001 John Wiley & Sons Ltd
ISBNs: 0-470-84137-0 (Hardback); 0-470-84534-1 (Electronic)
A random process may be defined as a set of random variables. We associate a time
function x(n) = x(n, A) with every possible outcome A of an experiment. Each time
function is called a realization of the random process or a random signal. The ensemble
of all these time functions (called sample functions) constitutes the random process x(n).
If we sample this process at some particular time n
0
, we obtain a random variable. Thus
a random process is a family of random variables.
We may consider the statistics of a random process in two ways. If we fix the time n at
n
0
and consider the random variable x(n
0
), we obtain statistics over the ensemble. For
example, E[x(n
0
)[ is the ensemble average, where E[[ is the expectation operation
introduced in Chapter 3. If we fix A and consider a particular sample function, we
have a time function and the statistics we obtain are temporal. For example, E[x(n, A
i
)[
is the time average. If the time average is equal to the ensemble average, we say that the
process is ergodic. The property of ergodicity is important because in practice we often
have access to only one sample function. Since we generally work only with temporal
statistics, it is important to be sure that the temporal statistics we obtain are the true
representation of the process as a whole.
8.1.1 Correlation Functions
For many applications, one signal is often used to compare with another in order to
determine the similarity between the pair, and to determine additional information
based on the similarity. Autocorrelation is used to quantify the similarity between two
segments of the same signal. The autocorrelation function of the random process x(n) is
defined as
r
xx
(n, k) = E[x(n)x(k)[: (8:1:1)
This function specifies the statistical relation of two samples at different time index n
and k, and gives the degree of dependence between two random variables of (n ÷k)
units apart. For example, consider a digital white noise x(n) as uncorrelated random
variables with zero-mean and variance s
2
x
. The autocorrelation function is
r
xx
(n, k) = E[x(n)x(k)[ = E[x(n)[E[x(k)[ =
0, n = k
s
2
x
, n = k.

(8:1:2)
If we subtract the means in (8.1.1) before taking the expected value, we have the
autocovariance function
g
xx
(n, k) = E¦[x(n) ÷m
x
(n)[[x(k) ÷m
x
(k)[¦ = r
xx
(n, k) ÷m
x
(n)m
x
(k): (8:1:3)
The objective in computing the correlation between two different random signals is
to measure the degree in which the two signals are similar. The crosscorrelation
and crosscovariance functions between two random processes x(n) and y(n) are defined
as
r
xy
(n, k) = E[x(n)y(k)[ (8:1:4)
352 ADAPTIVE FILTERING
and
g
xy
(n, k) = E¦[x(n) ÷m
x
(n)[[y(k) ÷m
y
(k)[¦ = r
xy
(n, k) ÷m
x
(n)m
y
(k): (8:1:5)
Correlation is a very useful DSP tool for detecting signals that are corrupted
by additive random noise, measuring the time delay between two signals, determining
the impulse response of a system (such as obtain the room impulse response used in
Section 4.5.2), and many others. Signal correlation is often used in radar, sonar, digital
communications, and other engineering areas. For example, in CDMA digital commu-
nications, data symbols are represented with a set of unique key sequences. If one of
these sequences is transmitted, the receiver compares the received signal with every
possible sequence from the set to determine which sequence has been received. In radar
and sonar applications, the received signal reflected from the target is the delayed
version of the transmitted signal. By measuring the round-trip delay, one can determine
the location of the target.
Both correlation functions and covariance functions are extensively used in analyzing
random processes. In general, the statistical properties of a random signal such as the
mean, variance, and autocorrelation and autocovariance functions are time-varying
functions. A random process is said to be stationary if its statistics do not change
with time. The most useful and relaxed form of stationary is the wide-sense stationary
(WSS) process. A random process is called WSS if the following two conditions are
satisfied:
1. The mean of the process is independent of time. That is,
E[x(n)[ = m
x
, (8:1:6)
1. where m
x
is a constant.
2. The autocorrelation function depends only on the time difference. That is,
r
xx
(k) = E[x(n ÷k)x(n)[: (8:1:7)
Equation (8.1.7) indicates that the autocorrelation function of a WSS process is inde-
pendent of the time shift and r
xx
(k) denotes the autocorrelation function of a time lag of
k samples.
The autocorrelation function r
xx
(k) of a WSS process has the following important
properties:
1. The autocorrelation function is an even function of the time lag k. That is,
r
xx
(÷k) = r
xx
(k): (8:1:8)
2. The autocorrelation function is bounded by the mean squared value of the process
expressed as
[r
xx
(k)[ _ r
xx
(0), (8:1:9)
INTRODUCTION TO RANDOM PROCESSES 353
2. where r
xx
(0) = E[x
2
(n)[ is equal to the mean-squared value, or the power in the
random process.
In addition, if x(n) is a zero-mean random process, we have
r
xx
(0) = E[x
2
(n)[ = s
2
x
: (8:1:10)
Thus the autocorrelation function of a signal has its maximum value at zero lag.
If x(n) has a periodic component, then r
xx
(k) will contain the same periodic com-
ponent.
Example 8.1: Given the sequence
x(n) = a
n
u(n), 0 < a < 1,
the autocorrelation function can be computed as
r
xx
(k) =
¸
·
n=÷·
x(n ÷k)x(n) =
¸
·
n=0
a
n÷k
a
n
= a
k
¸
·
n=0
(a
2
)
n
:
Since a < 0, we obtain
r
xx
(k) =
a
k
1 ÷a
2
:
Example 8.2: Consider the sinusoidal signal expressed as
x(n) = cos(!n),
find the mean and the autocorrelation function of x(n).
(a) m
x
= E[cos(!n)[ = 0.
(b) r
xx
(k) = E[x(n ÷k)x(n)[ = E[cos(!n ÷!k) cos(!n)[
=
1
2
E[cos(2!n ÷!k)[ ÷
1
2
cos(!k) =
1
2
cos(!k):
The crosscorrelation function of two WSS processes x(n) and y(n) is defined as
r
xy
(k) = E[x(n ÷k)y(n)[: (8:1:11)
This crosscorrelation function has the property
r
xy
(k) = r
yx
(÷k): (8:1:12)
Therefore r
yx
(k) is simply the folded version of r
xy
(k). Hence, r
yx
(k) provides exactly
the same information as r
xy
(k), with respect to the similarity of x(n) to y(n).
354 ADAPTIVE FILTERING
In practice, we only have one sample sequence ¦x(n)¦ available for analysis. As
discussed earlier, a stationary random process x(n) is ergodic if all its statistics can be
determined from a single realization of the process, provided that the realization is long
enough. Therefore time averages are equal to ensemble averages when the record length
is infinite. Since we do not have data of infinite length, the averages we compute differ
from the true values. In dealing with finite-duration sequence, the sample mean of x(n)
is defined as

m
x
=
1
N
¸
N÷1
n=0
x(n), (8:1:13)
where N is the number of samples in the short-time analysis interval. The sample
variance is defined as

s
2
x
=
1
N
¸
N÷1
n=0

x(n) ÷m
x

2
: (8:1:14)
The sample autocorrelation function is defined as

r
xx
(k) =
1
N ÷k
¸
N÷k÷1
n=0
x(n ÷k)x(n), k = 0, 1, . . . , N ÷1, (8:1:15)
where N is the length of the sequence x(n). Note that for a given sequence of length
N, Equation (8.1.15) generates values for up to N different lags. In practice, we can
only expect good results for lags of no more than 5±10 percent of the length of the
signals.
The autocorrelation and crosscorrelation functions introduced in this section can be
computed using the MATLAB function xcorr in the Signal Processing Toolbox. The
crosscorrelation function r
xy
(k) of the two sequences x(n) and y(n) can be computed
using the statement
c = xcorr(x, y);
where x and y are length N vectors and the crosscorrelation vector c has length 2N ÷1.
The autocorrelation function r
xx
(k) of the sequence x(n) can be computed using the
statement
c = xcorr(x);
In addition, the crosscovariance function can be estimated using
v = xcov(x, y);
and the autocovariance function can be computed with
v = xcov(x);
See Signal Processing Toolbox User's Guide for details.
INTRODUCTION TO RANDOM PROCESSES 355
8.1.2 Frequency-Domain Representations
In the study of deterministic digital signals, we use the discrete-time Fourier transform
(DTFT) or the z-transform to find the frequency contents of the signals. In this section,
we will use the same transform for random signals. Consider an ergodic random process
x(n). This sequence cannot be really representative of the random process because the
sequence x(n) is only one of infinitely possible sequences. However, if we consider the
autocorrelation function r
xx
(k), the result is always the same no matter which sample
sequence is used to compute r
xx
(k). Therefore we should apply the transform to r
xx
(k)
rather than x(n).
The correlation functions represent the time-domain description of the statistics of a
random process. The frequency-domain statistics are represented by the power density
spectrum (PDS) or the autopower spectrum. The PDS is the DTFT (or the z-transform)
of the autocorrelation function r
xx
(k) of a WSS signal x(n) defined as
P
xx
(!) =
¸
·
k=÷·
r
xx
(k)e
÷j!k
, (8:1:16)
or
P
xx
(z) =
¸
·
k=÷·
r
xx
(k)z
÷k
: (8:1:17)
A sufficient condition for the existence of the PDS is that r
xx
(k) is summable. The PDS
defined in (7.3.16) is equal to the DFT of the autocorrelation function. The windowing
technique introduced in Section 7.3.3 can be used to improve the convergence properties
of (7.3.16) and (7.3.17) if the DFT is used in computing the PDS of random signals.
Equation (8.1.16) implies that the autocorrelation function is the inverse DTFT of the
PDS, which is expressed as
r
xx
(k) =
1
2p

p
÷p
P
xx
(!)e
j!k
d!: (8:1:18)
From (8.1.10), we have the mean-square value
E[x
2
(n)[ = r
xx
(0) =
1
2p

p
÷p
P
xx
(!)d!: (8:1:19)
Thus r
xx
(0) represents the average power in the random signal x(n). The PDS is a
periodic function of the frequency !, with the period equal to 2p. We can show (in the
exercise problems) that P
xx
(!) of a WSS signal is a real-valued function of !. If x(n) is a
real-valued signal, P
xx
(!) is an even function of !. That is,
P
xx
(!) = P
xx
(÷!) (8:1:20)
or
P
xx
(z) = P
xx
(z
÷1
): (8:1:21)
356 ADAPTIVE FILTERING
The DTFT of the crosscorrelation function P
xy
(!) of two WSS signals x(n) and y(n) is
given by
P
xy
(!) =
¸
·
k=÷·
r
xy
(k)e
÷j!k
, (8:1:22)
or
P
xy
(z) =
¸
·
k=÷·
r
xy
(k)z
÷k
: (8:1:23)
This function is called the cross-power spectrum.
Example 8.3: The autocorrelation function of a WSS white random process can be
defined as
r
xx
(k) = s
2
x
d(k) ÷m
2
x
: (8:1:24)
The corresponding PDS is given by
P
xx
(!) = s
2
x
÷2pm
2
x
d(!), [![ _ p: (8:1:25)
An important white random signal is called white noise, which has zero mean.
Thus its autocorrelation function is expressed as
r
xx
(k) = s
2
x
d(k), (8:1:26)
and the power spectrum is given by
P
xx
(!) = s
2
x
, [![ < p, (8:1:27)
which is of constant value for all frequencies !.
Consider a linear and time-invariant digital filter defined by the impulse response
h(n), or the transfer function H(z). The input of the filter is a WSS random signal x(n)
with the PDS P
xx
(!). As illustrated in Figure 8.1, the PDS of the filter output y(n) can
be expressed as
P
yy
(!) =

H(!)

2
P
xx
(!) (8:1:28)
or
P
yy
(z) =

H(z)

2
P
xx
(z), (8:1:29)
INTRODUCTION TO RANDOM PROCESSES 357
h(n)
H(w)
x(n)
P
xx
(w) P
yy
(w)
y(n)
Figure 8.1 Linear filtering of random processes
where H(!) is the frequency response of the filter. Therefore the value of the output
PDS at frequency ! depends on the squared magnitude response of the filter and the
input PDS at the same frequency.
Another important relationships between x(n) and y(n) are
m
y
= E
¸
·
l=÷·
h(l)x(n ÷l)
¸ ¸
=
¸
·
l=÷·
h(l)E[x(n ÷l)[ = m
x
¸
·
l=÷·
h(l), (8:1:30)
and
r
yx
(k) = E[y(n ÷k)x(n)[ = E
¸
·
l=÷·
h(l)x(n ÷k ÷l)x(n)
¸ ¸
=
¸
·
l=÷·
h(l)r
xx
(k ÷l) = h(k) + r
xx
(k):
(8:1:31)
Taking the z-transform of both sides, we obtain
P
yx
(z) = H(z)P
xx
(z): (8:1:32)
Similarly, the relationships between the input and the output signals are
r
xy
(k) =
¸
·
l=÷·
h(l)r
xx
(k ÷l) = h(k) + r
xx
(÷k) (8:1:33)
and
P
yx
(z) = H
+
(z)P
xx
(z): (8:1:34)
If the input signal x(n) is a zero-mean white noise with the autocorrelation function
defined in (8.1.26), Equation (8.1.31) becomes
r
yx
(k) =
¸
·
l=÷·
h(l)s
2
x
d(k ÷l) = s
2
x
h(k): (8:1:35)
This equation shows that by computing the crosscorrelation function r
yx
(k), the impulse
response h(n) of a filter (or system) can be obtained. This fact can be used to estimate an
unknown system such as the room impulse response used in Chapter 4.
358 ADAPTIVE FILTERING
Example 8.4: Let the system shown in Figure 8.1 be a second-order FIR filter. The
input x(n) is a zero-mean white noise given by Example 8.3, and the I/O equation
is expressed as
y(n) = x(n) ÷3x(n ÷1) ÷2x(n ÷2):
Find the mean m
y
and the autocorrelation function r
yy
(k) of the output y(n).
(a) m
y
= E[ y(n)[ = E[x(n)[ ÷3E[x(n ÷1)[ ÷2E[x(n ÷2)[ = 0.
(b) r
yy
(k) = E[ y(n ÷k)y(n)[
= 14r
xx
(k) ÷9r
xx
(k ÷1) ÷9r
xx
(k ÷1) ÷2r
xx
(k ÷2) ÷2r
xx
(k ÷2)
=
14s
2
x
if k = 0
9s
2
x
if k = ±1
2s
2
x
if k = ±2
0 otherwise.

8.2 Adaptive Filters
Many practical applications involve the reduction of noise and distortion for extraction
of information from the received signal. The signal degradation in some physical
systems is time varying, unknown, or possibly both. Adaptive filters provide a useful
approach for these applications. Adaptive filters modify their characteristics to achieve
certain objectives and usually accomplish the modification (adaptation) automatically.
For example, consider a high-speed modem for transmitting and receiving data over
telephone channels. It employs a filter called a channel equalizer to compensate for
the channel distortion. Since the dial-up communication channels have different char-
acteristics on each connection and are time varying, the channel equalizers must be
adaptive.
Adaptive filters have received considerable attention from many researchers over the
past 30 years. Many adaptive filter structures and adaptation algorithms have been
developed for different applications. This chapter presents the most widely used adap-
tive filter based on the FIR filter with the LMS algorithm. Adaptive filters in this class
are relatively simple to design and implement. They are well understood with regard to
convergence speed, steady-state performance, and finite-precision effects.
8.2.1 Introduction to Adaptive Filtering
An adaptive filter consists of two distinct parts ± a digital filter to perform the desired
signal processing, and an adaptive algorithm to adjust the coefficients (or weights) of
that filter. A general form of adaptive filter is illustrated in Figure 8.2, where d(n) is a
desired signal (or primary input signal), y(n) is the output of a digital filter driven by a
reference input signal x(n), and an error signal e(n) is the difference between d(n) and
y(n). The function of the adaptive algorithm is to adjust the digital filter coefficients to
ADAPTIVE FILTERS 359
x(n) y(n)
d(n)
e(n)
+

Adaptive
algorithm
Digital
filter
Figure 8.2 Block diagram of adaptive filter
y(n)
z
−1
z
−1
w
0
(n)
x(n) x(n−1) x(n−L+1)
w
L−1
(n) w
1
(n)
Figure 8.3 Block diagram of FIR filter for adaptive filtering
minimize the mean-square value of e(n). Therefore the filter weights are updated so that
the error is progressively minimized on a sample-by-sample basis.
In general, there are two types of digital filters that can be used for adaptive filtering:
FIR and IIR filters. The choice of an FIR or an IIR filter is determined by practical
considerations. The FIR filter is always stable and can provide a linear phase response.
On the other hand, the IIR filter involves both zeros and poles. Unless they are properly
controlled, the poles in the filter may move outside the unit circle and make the filter
unstable. Because the filter is required to be adaptive, the stability problems are much
difficult to handle. Thus the FIR adaptive filter is widely used for real-time applications.
The discussions in the following sections will be restricted to the class of adaptive FIR
filters.
The most widely used adaptive FIR filter is depicted in Figure 8.3. Given a set
of L coefficients, w
l
(n), l = 0, 1, . . . , L ÷1, and a data sequence, ¦x(n) x(n ÷1)
. . . x(n ÷L ÷1)¦, the filter output signal is computed as
y(n) =
¸
L÷1
l=0
w
l
(n)x(n ÷l), (8:2:1)
where the filter coefficients w
l
(n) are time varying and updated by the adaptive algo-
rithms that will be discussed next.
We define the input vector at time n as
x(n) = [x(n) x(n ÷1) . . . x(n ÷L ÷1)[
T
(8:2:2)
360 ADAPTIVE FILTERING
and the weight vector at time n as
w(n) = [w
0
(n) w
1
(n) . . . w
L÷1
(n)[
T
: (8:2:3)
Then the output signal y(n) in (8.2.1) can be expressed using the vector operation
y(n) = w
T
(n)x(n) = x
T
(n)w(n): (8:2:4)
The filter output y(n) is compared with the desired response d(n), which results in the
error signal
e(n) = d(n) ÷y(n) = d(n) ÷w
T
(n)x(n): (8:2:5)
In the following sections, we assume that d(n) and x(n) are stationary, and our objective is
to determine the weight vector so that the performance (or cost) function is minimized.
8.2.2 Performance Function
The general block diagram of the adaptive filter shown in Figure 8.2 updates the
coefficients of the digital filter to optimize some predetermined performance criterion.
The most commonly used performance measurement is based on the mean-square error
(MSE) defined as
x(n) = E[e
2
(n)[: (8:2:6)
For an adaptive FIR filter, x(n) will depend on the L filter weights w
0
(n), w
1
(n),
. . . , w
L÷1
(n). The MSE function can be determined by substituting (8.2.5) into (8.2.6),
expressed as
x(n) = E[d
2
(n)[ ÷2p
T
w(n) ÷w
T
(n)Rw(n), (8:2:7)
where p is the crosscorrelation vector defined as
p = E[d(n)x(n)[ = [r
dx
(0) r
dx
(1) . . . r
dx
(L ÷1)[
T
, (8:2:8)
and
r
dx
(k) = E[d(n)x(n ÷k)[ (8:2:9)
is the crosscorrelation function between d(n) and x(n). In (8.2.7), R is the input auto-
correlation matrix defined as
R = E[x(n)x
T
(n)[ =
r
xx
(0) r
xx
(1) r
xx
(L ÷1)
r
xx
(1) r
xx
(0) r
xx
(L ÷2)
.
.
.

.
.
.
.
.
.
r
xx
(L ÷1) r
xx
(L ÷2) r
xx
(0)

¸
¸
¸
¸
¸
, (8:2:10)
ADAPTIVE FILTERS 361
where
r
xx
(k) = E[x(n)x(n ÷k)[ (8:2:11)
is the autocorrelation function of x(n).
Example 8.5: Given an optimum filter illustrated in the following figure:
x(n) x(n−1)
w
1
z
−1
+ +
+

d(n)
e(n)
If E[x
2
(n)[ = 1, E[x(n)x(n ÷1)[ = 0:5, E[d
2
(n)[ = 4, E[d(n)x(n)[ = ÷1, and
E[d(n)x(n ÷1)[ = 1. Find x.
From (8.2.10), R =
1 0:5
0:5 1
¸
, and from (8.2.8), we have P =
÷1
1
¸
.
Therefore from (8.2.7), we obtain
x = E[d
2
(n)[ ÷2p
T
w ÷w
T
Rw
= 4 ÷2[÷1 1[
1
w
1
¸
÷[1 w
1
[
1 0:5
0:5 1
¸
1
w
1
¸
= w
2
1
÷w
1
÷7:
The optimum filter w
o
minimizes the MSE cost function x(n). Vector differentiation
of (8.2.7) gives w
o
as the solution to
Rw
o
= p: (8:2:12)
This system equation defines the optimum filter coefficients in terms of two correlation
functions ± the autocorrelation function of the filter input and the crosscorrelation
function between the filter input and the desired response. Equation (8.2.12) provides a
solution to the adaptive filtering problem in principle. However, in many applications,
the signal may be non-stationary. This linear algebraic solution, w
o
= R
÷1
p, requires
continuous estimation of R and p, a considerable amount of computations. In addition,
when the dimension of the autocorrelation matrix is large, the calculation of R
÷1
may
present a significant computational burden. Therefore a more useful algorithm is
obtained by developing a recursive method for computing w
o
, which will be discussed
in the next section.
To obtain the minimum MSE, we substitute the optimum weight vector w
o
= R
÷1
p
for w(n) in (8.2.7), resulting in
x
min
= E[d
2
(n)[ ÷p
T
w
o
: (8:2:13)
362 ADAPTIVE FILTERING
Since R is positive semidefinite, the quadratic form on the right-hand side of (8.2.7)
indicates that any departure of the weight vector w(n) from the optimum w
o
would
increase the error above its minimum value. In other words, the error surface is concave
and possesses a unique minimum. This feature is very useful when we utilize search
techniques in seeking the optimum weight vector. In such cases, our objective is to
develop an algorithm that can automatically search the error surface to find the
optimum weights that minimize x(n) using the input signal x(n) and the error signal e(n).
Example 8.6: Consider a second-order FIR filter with two coefficients w
0
and
w
1
, the desired signal d(n) =

2

sin(n!
0
), n _ 0, and the reference signal
x(n) = d(n ÷1). Find w
o
and x
min
Similar to Example 8.2, we can obtain r
xx
(0) = E[x
2
(n)[ = E[d
2
(n)[ = 1,
r
xx
(1) = cos(!
0
), r
xx
(2) = cos(2!
0
), r
dx
(0) = r
xx
(1), and r
dx
(1) = r
xx
(2). From
(8.2.12), we have
w
o
= R
÷1
p =
1 cos(!
0
)
cos(!
0
) 1
¸
÷1
cos(!
o
)
cos(2!
0
)
¸
=
2 cos(!
0
)
÷1
¸
:
From (8.2.13), we obtain
x
min
= 1 ÷[cos(!
0
) cos(2!
0
)[
2 cos(!
0
)
÷1
¸
= 0:
Equation (8.2.7) is the general expression for the performance function of an adaptive
FIR filter with given weights. That is, the MSE is a function of the filter coefficient
vector w(n). It is important to note that the MSE is a quadratic function because the
weights appear only to the first and second degrees in (8.2.7). For each coefficient vector
w(n), there is a corresponding (scalar) value of MSE. Therefore the MSE values
associated with w(n) form an (L ÷1)-dimensional space, which is commonly called the
MSE surface, or the performance surface.
For L = 2, this corresponds to an error surface in a three-dimensional space. The
height of x(n) corresponds to the power of the error signal e(n) that results from filtering
the signal x(n) with the coefficients w(n). If the filter coefficients change, the power in the
error signal will also change. This is indicated by the changing height on the surface
above w
0
÷w
1
the plane as the component values of w(n) are varied. Since the error
surface is quadratic, a unique filter setting w(n) = w
o
will produce the minimum MSE,
x
min
. In this two-weight case, the error surface is an elliptic paraboloid. If we cut the
paraboloid with planes parallel to the w
0
÷w
1
plane, we obtain concentric ellipses of
constant mean-square error. These ellipses are called the error contours of the error
surface.
Example 8.7: Consider a second-order FIR filter with two coefficients w
0
and w
1
.
The reference signal x(n) is a zero-mean white noise with unit variance. The
desired signal is given as
d(n) = b
0
x(n) ÷b
1
x(n ÷1):
ADAPTIVE FILTERS 363
Plot the error surface and error contours.
From Equation (8.2.10), we obtain R =
r
xx
(0) r
xx
(1)
r
xx
(1) r
xx
(0)
¸
=
1 0
0 1
¸
. From
(8.2.8), we have p =
r
dx
(0)
r
dx
(1)
¸
=
b
0
b
1
¸
. From (8.2.7), we get
x = E[d
2
(n)[ ÷2p
T
w ÷w
T
Rw = (b
2
0
÷b
2
1
) ÷2b
0
w
0
÷2b
1
w
1
÷w
2
0
÷w
2
1
Let b
0
= 0:3 and b
1
= 0:5, we have
x = 0:34 ÷0:6w
0
÷w
1
÷w
2
0
÷w
2
1
:
The MATLAB script (exam8_7a.m in the software package) is used to plot the error
surface shown in Figure 8.4(a) and the script exam8_7b.m is used to plot the error
contours shown in Figure 8.4(b).
1200
1000
800
600
M
S
E
400
200
20
15
10
5
0
−5
−10
−15
−20
−30 −20 −10 0
w0
Error Contour
Error Surface
w
1
10 20 30
0
40
20
0
−20
−40
40
20
0
w0 w1
−20
−40
Figure 8.4 Performance surface and error contours, L = 2
364 ADAPTIVE FILTERING
One of the most important properties of the MSE surface is that it has only one global
minimum point. At that minimum point, the tangents to the surface must be 0. Minim-
izing the MSE is the objective of many current adaptive methods such as the LMS
algorithm.
8.2.3 Method of Steepest Descent
As shown in Figure 8.4, the MSE of (8.2.7) is a quadratic function of the weights that
can be pictured as a positive-concave hyperparabolic surface. Adjusting the weights to
minimize the error involves descending along this surface until reaching the `bottom of
the bowl.' Various gradient-based algorithms are available. These algorithms are based
on making local estimates of the gradient and moving downward toward the bottom of
the bowl. The selection of an algorithm is usually decided by the speed of convergence,
steady-state performance, and the computational complexity.
The steepest-descent method reaches the minimum by following the direction in
which the performance surface has the greatest rate of decrease. Specifically, an algo-
rithm whose path follows the negative gradient of the performance surface. The
steepest-descent method is an iterative (recursive) technique that starts from some initial
(arbitrary) weight vector. It improves with the increased number of iterations. Geomet-
rically, it is easy to see that with successive corrections of the weight vector in the
direction of the steepest descent on the concave performance surface, we should arrive
at its minimum, x
min
, at which point the weight vector components take on their
optimum values. Let x(0) represent the value of the MSE at time n = 0 with an arbitrary
choice of the weight vector w(0). The steepest-descent technique enables us to descend
to the bottom of the bowl, w
o
, in a systematic way. The idea is to move on the error
surface in the direction of the tangent at that point. The weights of the filter are updated
at each iteration in the direction of the negative gradient of the error surface.
The mathematical development of the method of steepest descent is easily seen from
the viewpoint of a geometric approach using the MSE surface. Each selection of a filter
weight vector w(n) corresponds to only one point on the MSE surface, [w(n), x(n)].
Suppose that an initial filter setting w(0) on the MSE surface, [w(0), x(0)] is arbitrarily
chosen. A specific orientation to the surface is then described using the directional
derivatives of the surface at that point. These directional derivatives quantify the rate of
change of the MSE surface with respect to the w(n) coordinate axes. The gradient of the
error surface \x(n) is defined as the vector of these directional derivatives.
The concept of steepest descent can be implemented in the following algorithm:
w(n ÷1) = w(n) ÷
m
2
\x(n) (8:2:14)
where m is a convergence factor (or step size) that controls stability and the rate of
descent to the bottom of the bowl. The larger the value of m, the faster the speed of
descent. The vector \x(n) denotes the gradient of the error function with respect to
w(n), and the negative sign increments the adaptive weight vector in the negative
gradient direction. The successive corrections to the weight vector in the direction of
ADAPTIVE FILTERS 365
the steepest descent of the performance surface should eventually lead to the minimum
mean-square error x
min
, at which point the weight vector reaches its optimum value w
o
.
When w(n) has converged to w
o
, that is, when it reaches the minimum point of the
performance surface, the gradient \x(n) = 0. At this time, the adaptation in (8.2.14) is
stopped and the weight vector stays at its optimum solution. The convergence can be
viewed as a ball placed on the `bowl-shaped' MSE surface at the point [w(0), x(0)]. If the
ball was released, it would roll toward the minimum of the surface, and would initially
roll in a direction opposite to the direction of the gradient, which can be interpreted as
rolling towards the bottom of the bowl.
8.2.4 The LMS Algorithm
From (8.2.14), we see that the increment from w(n) to w(n ÷1) is in the negative
gradient direction, so the weight tracking will closely follow the steepest descent path
on the performance surface. However, in many practical applications the statistics of
d(n) and x(n) are unknown. Therefore the method of steepest descent cannot be used
directly, since it assumes exact knowledge of the gradient vector at each iteration.
Widrow [13] used the instantaneous squared error, e
2
(n), to estimate the MSE. That is,
^
x(n) = e
2
(n): (8:2:15)
Therefore the gradient estimate used by the LMS algorithm is
\
^
x(n) = 2

\e(n)

e(n): (8:2:16)
Since e(n) = d(n) ÷w
T
(n)x(x), \e(n) = ÷x(n), the gradient estimate becomes
\
^
x(n) = ÷2x(n)e(n): (8:2:17)
Substituting this gradient estimate into the steepest-descent algorithmof (8.2.14), we have
w(n ÷1) = w(n) ÷mx(n)e(n): (8:2:18)
This is the well-known LMS algorithm, or stochastic gradient algorithm. This algorithm
is simple and does not require squaring, averaging, or differentiating. The LMS algo-
rithm provides an alternative method for determining the optimum filter coefficients
without explicitly computing the matrix inversion suggested in (8.2.12).
Widrow's LMS algorithm is illustrated in Figure 8.5 and is summarized as follows:
1. Determine L, m, and w(0), where L is the order of the filter, m is the step size, and
w(0) is the initial weight vector at time n = 0.
2. Compute the adaptive filter output
y(n) =
¸
L÷1
l=0
w
l
(n)x(n ÷l): (8:2:19)
366 ADAPTIVE FILTERING
x(n) y(n)
d(n)
e(n)
+

w(n)
LMS
Figure 8.5 Block diagram of an adaptive filter with the LMS algorithm
3. Compute the error signal
e(n) = d(n) ÷y(n): (8:2:20)
4. Update the adaptive weight vector from w(n) to w(n + 1) by using the LMS
algorithm
w
l
(n ÷1) = w
l
(n) ÷mx(n ÷l)e(n), l = 0, 1, . . . , L ÷1: (8:2:21)
8.3 Performance Analysis
A detailed discussion of the performance of the LMS algorithm is available in many
textbooks. In this section, we present some important properties of the LMS algorithm
such as stability, convergence rate, and the excess mean-square error due to gradient
estimation error.
8.3.1 Stability Constraint
As shown in Figure 8.5, the LMS algorithm involves the presence of feedback. Thus
the algorithm is subject to the possibility of becoming unstable. From (8.2.18), we
observe that the parameter m controls the size of the incremental correction applied
to the weight vector as we adapt from one iteration to the next. The mean weight
convergence of the LMS algorithm from initial condition w(0) to the optimum filter w
o
must satisfy
0 < m <
2
l
max
, (8:3:1)
where l
max
is the largest eigenvalue of the autocorrelation matrix R defined in (8.2.10).
Applying the stability constraint on m given in (8.3.1) is difficult because of the compu-
tation of l
max
when L is large.
In practical applications, it is desirable to estimate l
max
using a simple method. From
(8.2.10), we have
PERFORMANCE ANALYSIS 367
tr[R[ = Lr
xx
(0) =
¸
L÷1
l=0
l
l
, (8:3:2)
where tr[R] denotes the trace of matrix R. It follows that
l
max
_
¸
L÷1
l=0
l
l
= Lr
xx
(0) = LP
x
, (8:3:3)
where
P
x
= r
xx
(0) = E

x
2
(n)

(8:3:4)
denotes the power of x(n). Therefore setting
0 < m <
2
LP
x
(8:3:5)
assures that (8.3.1) is satisfied.
Equation (8.3.5) provides some important information on how to select m, and they
are summarized as follows:
1. Since the upper bound on m is inversely proportional to L, a small m is used for large-
order filters.
2. Since m is made inversely proportional to the input signal power, weaker signals use
a larger m and stronger signals use a smaller m. One useful approach is to normalize
with respect to the input signal power P
x
. The resulting algorithm is called the
normalized LMS algorithm, which will be discussed in Section 8.4.
8.3.2 Convergence Speed
In the previous section, we saw that w(n) converges to w
o
if the selection of m satisfies
(8.3.1). Convergence of the weight vector w(n) from w(0) to w
o
corresponds to the
convergence of the MSE from x(0) to x
min
. Therefore convergence of the MSE toward
its minimum value is a commonly used performance measurement in adaptive systems
because of its simplicity. During adaptation, the squared error e
2
(n) is non-stationary as
the weight vector w(n) adapts toward w
o
. The corresponding MSE can thus be defined
only based on ensemble averages. A plot of the MSE versus time n is referred to as the
learning curve for a given adaptive algorithm. Since the MSE is the performance
criterion of LMS algorithms, the learning curve is a natural way to describe the transient
behavior.
Each adaptive mode has its own time constant, which is determined by the overall
adaptation constant m and the eigenvalue l
l
associated with that mode. Overall con-
vergence is clearly limited by the slowest mode. Thus the overall MSE time constant can
be approximated as
368 ADAPTIVE FILTERING
t
mse
÷
1
ml
min
, (8:3:6)
where l
min
is the minimum eigenvalue of the R matrix. Because t
mse
is inversely propor-
tional to m, we have a large t
mse
when m is small (i.e., the speed of convergence is slow). If
we use a large value of m, the time constant is small, which implies faster convergence.
The maximum time constant t
mse
= 1=ml
min
is a conservative estimate of filter per-
formance, since only large eigenvalues will exert significant influence on the conver-
gence time. Since some of the projections may be negligibly small, the adaptive filter
error convergence may be controlled by fewer modes than the number of adaptive
filter weights. Consequently, the MSE often converges more rapidly than the upper
bound of (8.3.6) would suggest.
Because the upper bound of t
mse
is inversely proportional to l
min
, a small l
min
can
result in a large time constant (i.e., a slow convergence rate). Unfortunately, if l
max
is
also very large, the selection of m will be limited by (8.3.1) such that only a small m can
satisfy the stability constraint. Therefore if l
max
is very large and l
min
is very small, from
(8.3.6), the time constant can be very large, resulting in very slow convergence. As
previously noted, the fastest convergence of the dominant mode occurs for m = 1=l
max
.
Substituting this smallest step size into (8.3.6) results in
t
mse
_
l
max
l
min
: (8:3:7)
For stationary input and sufficiently small m, the speed of convergence of the algorithm
is dependent on the eigenvalue spread (the ratio of the maximum to minimum eigen-
values) of the matrix R.
As mentioned in the previous section, the eigenvalues l
max
and l
min
are very difficult
to compute. However, there is an efficient way to estimate the eigenvalue spread from
the spectral dynamic range. That is,
l
max
l
min
_
max [X(!)[
2
min [X(!)[
2
, (8:3:8)
where X(!) is DTFT of x(n) and the maximum and minimum are calculated over the
frequency range 0 _ ! _ p. From (8.3.7) and (8.3.8), input signals with a flat (white)
spectrum have the fastest convergence speed.
8.3.3 Excess Mean-Square Error
The steepest-descent algorithm in (8.2.14) requires knowledge of the gradient \x(n),
which must be estimated at each iteration. The estimated gradient \
^
x(n) produces the
gradient estimation noise. After the algorithm converges, i.e., w(n) is close to w
o
, the
true gradient \x(n) ~ 0. However, the gradient estimator \
^
x(n) = 0. As indicated by
the update of Equation (8.2.14), perturbing the gradient will cause the weight vector
w(n ÷1) to move away from the optimum solution w
o
. Thus the gradient estimation
PERFORMANCE ANALYSIS 369
noise prevents w(n ÷1) from staying at w
o
in steady state. The result is that w(n) varies
randomly about w
o
. Because w
o
corresponds to the minimum MSE, when w(n) moves
away from w
o
, it causes x(n) to be larger than its minimum value, x
min
, thus producing
excess noise at the filter output.
The excess MSE, which is caused by random noise in the weight vector after con-
vergence, is defined as the average increase of the MSE. For the LMS algorithm, it can
be approximated as
x
excess
~
m
2
LP
x
x
min
: (8:3:9)
This approximation shows that the excess MSE is directly proportional to m. The larger
the value of m, the worse the steady-state performance after convergence. However,
Equation (8.3.6) shows that a larger m results in faster convergence. There is a design
trade-off between the excess MSE and the speed of convergence.
The optimal step size m is difficult to determine. Improper selection of m might make
the convergence speed unnecessarily slow or introduce excess MSE. If the signal is non-
stationary and real-time tracking capability is crucial for a given application, then use a
larger m. If the signal is stationary and convergence speed is not important, use a smaller
m to achieve better performance in a steady state. In some practical applications, we can
use a larger m at the beginning of the operation for faster convergence, then use a smaller
m to achieve better steady-state performance.
The excess MSE, x
excess
, in (8.3.9) is also proportional to the filter order L, which
means that a larger L results in larger algorithm noise. From (8.3.5), a larger L implies a
smaller m, resulting in slower convergence. On the other hand, a large L also implies
better filter characteristics such as sharp cutoff. There exists an optimum order L for
any given application. The selection of L and m also will affect the finite-precision error,
which will be discussed in Section 8.6.
In a stationary environment, the signal statistics are unknown but fixed. The LMS
algorithm gradually learns the required input statistics. After convergence to a steady
state, the filter weights jitter around the desired fixed values. The algorithm perform-
ance is determined by both the speed of convergence and the weight fluctuations in
steady state. In the non-stationary case, the algorithm must continuously track the time-
varying statistics of the input. Performance is more difficult to assess.
8.4 Modified LMS Algorithms
The LMS algorithm described in the previous section is the most widely used adaptive
algorithm for practical applications. In this section, we present two modified algorithms
that are the direct variants of the basic LMS algorithm.
8.4.1 Normalized LMS Algorithm
The stability, convergence speed, and fluctuation of the LMS algorithm are governed by
the step size m and the reference signal power. As shown in (8.3.5), the maximum stable
370 ADAPTIVE FILTERING
step-size m is inversely proportional to the filter order L and the power of the reference
signal x(n). One important technique to optimize the speed of convergence while
maintaining the desired steady-state performance, independent of the reference signal
power, is known as the normalized LMS algorithm (NLMS). The NLMS algorithm is
expressed as
w(n ÷1) = w(n) ÷m(n)x(n)e(n), (8:4:1)
where m(n) is an adaptive step size that is computed as
m(n) =
a
L
^
P
x
(n),
(8:4:2)
where
^
P
x
(n) is an estimate of the power of x(n) at time n, and a is a normalized step size
that satisfies the criterion
0 < a < 2: (8:4:3)
The commonly used method to estimate
^
P
x
(n) sample-by-sample is introduced in
Section 3.2.1. Some useful implementation considerations are given as follows:
1. Choose
^
P
x
(0) as the best a priori estimate of the reference signal power.
2. Since it is not desirable that the power estimate
^
P
x
(n) be zero or very small, a
software constraint is required to ensure that m(n) is bounded even if
^
P
x
(n) is very
small when the signal is absent for a long time. This can be achieved by modifying
(8.4.2) as
m(n) =
a
L
^
P
x
(n) ÷c,
(8:4:4)
2. where c is a small constant.
8.4.2 Leaky LMS Algorithm
Insufficient spectral excitation of the LMS algorithm may result in divergence of the
adaptive weights. In that case, the solution is not unique and finite-precision effects can
cause the unconstrained weights to grow without bound, resulting in overflow during the
weight update process. This long-terminstability is undesirable for real-time applications.
Divergence can often be avoided by means of a `leaking' mechanism used during the
weight update calculation. This is called the leaky LMS algorithm and is expressed as
w(n ÷1) = vw(n) ÷mx(n)e(n), (8:4:5)
where v is the leakage factor with 0 < v _ 1. It can be shown that leakage is the
deterministic equivalent of adding low-level white noise. Therefore this approach results
MODIFIED LMS ALGORITHMS 371
in some degradation in adaptive filter performance. The value of the leakage factor is
determined by the designer on an experimental basis as a compromise between robust-
ness and loss of performance of the adaptive filter. The leakage factor introduces a bias
on the long-term coefficient estimation. The excess error power due to the leakage is
proportional to [(1 ÷v)=m[
2
. Therefore (1 ÷v) should be kept smaller than m in order to
maintain an acceptable level of performance. For fixed-point hardware realization,
multiplication of each coefficient by v, as shown in (8.4.5), can lead to the introduction
of roundoff noise, which adds to the excess MSE. Therefore the leakage effects must be
incorporated into the design procedure for determining the required coefficient and
internal data wordlength. The leaky LMS algorithm not only prevents unconstrained
weight overflow, but also limits the output power in order to avoid nonlinear distortion.
8.5 Applications
The desirable features of an adaptive filter are the ability to operate in an unknown
environment and to track time variations of the input signals, making it a powerful
algorithm for DSP applications. The essential difference between various applications
of adaptive filtering is where the signals x(n), d(n), y(n), and e(n) are connected. There
are four basic classes of adaptive filtering applications: identification, inverse modeling,
prediction, and interference canceling.
8.5.1 Adaptive System Identification
Systemidentification is an experimental approach to the modeling of a process or a plant.
The basic idea is to measure the signals produced by the system and to use them to
construct a model. The paradigm of system identification is illustrated in Figure 8.6,
where P(z) is an unknown systemto be identified and W(z) is a digital filter used to model
P(z). By exciting both the unknown system P(z) and the digital model W(z) with the same
excitation signal x(n) and measuring the output signals y(n) and d(n), we can determine
the characteristics of P(z) by adjusting the digital model W(z) to minimize the difference
between these two outputs. The digital model W(z) can be an FIR filter or an IIR filter.
x(n) y(n)
d(n)
e(n)
+
− Digital
filter, W(z)
LMS
algorithm
Unknown
system, P(z)
Signal
generator
Figure 8.6 Block diagram of adaptive system identification using the LMS algorithm
372 ADAPTIVE FILTERING
Adaptive system identification is a technique that uses an adaptive filter for the model
W(z). This section presents the application of adaptive estimation techniques for direct
system modeling. This technique has been widely applied in echo cancellation, which
will be introduced in Sections 9.4 and 9.5. A further application for system modeling is
to estimate various transfer functions in active noise control systems [8].
Adaptive system identification is a very important procedure that is used frequently
in the fields of control systems, communications, and signal processing. The modeling
of a single-input/single-output dynamic system (or plant) is shown in Figure 8.6, where
x(n), which is usually white noise, is applied simultaneously to the adaptive filter and the
unknown system. The output of the unknown system then becomes the desired signal,
d(n), for the adaptive filter. If the input signal x(n) provides sufficient spectral excita-
tion, the adaptive filter output y(n) will approximate d(n) in an optimum sense after
convergence.
Identification could mean that a set of data is collected from the system, and that a
separate procedure is used to construct a model. Such a procedure is usually called off-
line (or batch) identification. In many practical applications, however, the model is
sometimes needed on-line during the operation of the system. That is, it is necessary to
identify the model at the same time that the data set is collected. The model is updated at
each time instant that a new data set becomes available. The updating is performed with
a recursive adaptive algorithm such as the LMS algorithm.
As shown in Figure 8.6, it is desired to learn the structure of the unknown system
from knowledge of its input x(n) and output d(n). If the unknown time-invariant system
P(z) can be modeled using an FIR filter of order L, the estimation error is given as
e(n) = d(n) ÷y(n) =
¸
L÷1
l=0

p(l) ÷w
l
(n)

x(n ÷l), (8:5:1)
where p(l) is the impulse response of the unknown plant.
By choosing each w
l
(n) close to each p(l), the error will be made small. For white-
noise input, the converse also holds: minimizing e(n) will force the w
l
(n) to approach
p(l), thus identifying the system
w
l
(n) ~ p(l), l = 0, 1, . . . , L ÷1: (8:5:2)
The basic concept is that the adaptive filter adjusts itself, intending tocause its output to
match that of the unknown system. When the difference between the physical system
response d(n) and adaptive model response y(n) has been minimized, the adaptive model
approximates P(z). In actual applications, there will be additive noise present at the
adaptive filter input and so the filter structure will not exactly match that of the unknown
system. When the plant is time varying, the adaptive algorithmhas the task of keeping the
modeling error small by continually tracking time variations of the plant dynamics.
8.5.2 Adaptive Linear Prediction
Linear prediction is a classic signal processing technique that provides an estimate of the
value of an input process at a future time where no measured data is yet available. The
APPLICATIONS 373
techniques have been successfully applied to a wide range of applications such as speech
coding and separating signals from noise. As illustrated in Figure 8.7, the time-domain
predictor consists of a linear prediction filter in which the coefficients w
l
(n) are updated
with the LMS algorithm. The predictor output y(n) is expressed as
y(n) =
¸
L÷1
l=0
w
l
(n)x(n ÷D ÷l), (8:5:3)
where the delay D is the number of samples involved in the prediction distance of the
filter. The coefficients are updated as
w(n ÷1) = w(n) ÷mx(n ÷D)e(n), (8:5:4)
where x(n ÷D) = [x(n ÷D) x(n ÷D ÷1) . . . x(n ÷D ÷L ÷1)[
T
is the delayed refer-
ence signal vector, and e(n) = x(n) ÷y(n) is the prediction error. Proper selection of the
prediction delay D allows improved frequency estimation performance for multiple
sinusoids in white noise.
Now consider the adaptive predictor for enhancing an input of M sinusoids
embedded in white noise, which is of the form
x(n) = s(n) ÷v(n) =
¸
M÷1
m=0
A
m
sin(!
m
n ÷f
m
) ÷v(n), (8:5:5)
where v(n) is white noise with uniform noise power s
2
v
. In this application, the structure
shown in Figure 8.7 is called the adaptive line enhancer, which provides an efficient
means for the adaptive tracking of the sinusoidal components of a received signal x(n)
and separates these narrowband signals s(n) from broadband noise v(n). This technique
has been shown effective in practical applications when there is insufficient a priori
knowledge of the signal and noise parameters.
As shown in Figure 8.7, we want the highly correlated components of x(n) to appear
in y(n). This is accomplished by adjusting the weights to minimize the expected mean-
square value of the error signal e(n). This causes an adaptive filter W(z) to form
x(n)
y(n) e(n)
+

Digital
filter W(z)
LMS
z
−∆
y(n)
Broadband
output
Narrowband
output
Figure 8.7 Block diagram of an adaptive predictor
374 ADAPTIVE FILTERING
narrowband bandpass filters centered at the frequency of the sinusoidal components.
The noise component of the input is rejected, while the phase difference (caused by D) of
the narrowband signals is readjusted so that they can cancel correlated components in
d(n) to minimize the error signal e(n). In this case, the output y(n) is the enhanced signal,
which contains multiple sinusoidals as expressed in (8.5.5).
In many digital communications and signal detection applications, the desired broad-
band (spread-spectrum) signal v(n) is corrupted by an additive narrowband interference
s(n). From a filtering viewpoint, the objective of an adaptive filter is to form a notch in
the frequency band occupied by the interference, thus suppressing the narrowband
noise. The narrowband characteristics of the interference allow W(z) to estimate s(n)
from past samples of x(n) and to subtract the estimate from x(n). The error signal e(n) in
Figure 8.7 consists of desired broadband signals. In this application, the desired output
from the overall interference suppression filter is e(n).
8.5.3 Adaptive Noise Cancellation
The wide spread use of cellular phones has significantly increased the use of com-
munication systems in high noise environments. Intense background noise, however,
often corrupts speech and degrading the performance of many communication
systems. Existing signal processing techniques such as speech coding, automatic speech
recognition, speaker identification, channel transmission, and echo cancellation are
developed under noise-free assumptions. These techniques could be employed in noisy
environments if a front-end noise suppression algorithm sufficiently reduces additive
noise. Noise reduction is becoming increasingly important with the development and
application of hands-free and voice-activated cellular phones.
Single-channel noise reduction methods involve Wiener filtering, Kalman filtering,
and spectral subtraction. In the dual-channel systems, a second sensor provides a
reference noise to better characterize changing noise statistics, which is necessary for
dealing with non-stationary noise. The most widely used dual-channel adaptive noise
canceler (ANC) employs an adaptive filter with the LMS algorithm to cancel the noise
component in the primary signal picked up by the primary sensor.
As illustrated in Figure 8.8, the basic concept of adaptive noise cancellation is to
process signals from two sensors and to reduce the level of the undesired noise with
adaptive filtering techniques. The primary sensor is placed close to the signal source in
order to pick up the desired signal. However, the primary sensor output also contains
noise from the noise source. The reference sensor is placed close to the noise source
to sense only the noise. This structure takes advantage of the correlation between the
noise signals picked up by the primary sensor and those picked up by the reference
sensor.
A block diagram of the adaptive noise cancellation system is illustrated in Figure 8.9,
where P(z) represents the transfer function between the noise source and the primary
sensor. The canceler has two inputs: the primary input d(n) and the reference input x(n).
The primary input d(n) consists of signal s(n) plus noise x
/
(n), i.e., d(n) = s(n) ÷x
/
(n),
which is highly correlated with x(n) since they are derived from the same noise source.
The reference input simply consists of noise x(n). The objective of the adaptive filter is to
use the reference input x(n) to estimate the noise x
/
(n). The filter output y(n), which is an
APPLICATIONS 375

+
Noise
source
y(n)
Signal
source
LMS
Primary
sensor
Reference
sensor
d(n)
x(n)
W(z)
e(n)
z
−1
Figure 8.8 Basic concept of adaptive noise canceling
x(n) y(n)
d(n)
e(n)
+
+
+
− Digital
filter
Adaptive
algorithm
s(n)
x′(n)
P(z)
Figure 8.9 Block diagram of adaptive noise canceler
estimate of noise x
/
(n), is then subtracted from the primary channel signal d(n), produ-
cing e(n) as the desired signal plus reduced noise.
To minimize the residual error e(n), the adaptive filter W(z) will generate an output
y(n) that is an approximation of x
/
(n). Therefore the adaptive filter W(z) will converge
to the unknown plant P(z). This is the adaptive system identification scheme discussed
in Section 8.5.1. To apply the ANC effectively, the reference noise picked up by the
reference sensor must be highly correlated with the noise components in the primary
signal. This condition requires a close spacing between the primary and reference
sensors. Unfortunately, it is also critical to avoid the signal components from the signal
source being picked up by the reference sensor. This `crosstalk' effect will degrade the
performance of ANC because the presence of the signal components in reference signal
will cause the ANC to cancel the desired signal along with the undesired noise. The
performance degradation of ANC with crosstalk includes less noise reduction, slower
convergence, and reverberant distortion in the desired signal.
Crosstalk problems may be eliminated by placing the primary sensor far away from
the reference sensor. Unfortunately, this arrangement requires a large-order filter in
order to obtain adequate noise reduction. For example, a separation of a few meters
between the two sensors requires a filter with 1500 taps to achieve 20 dB noise reduction.
The long filter increases excess mean-square error and decreases the tracking ability of
376 ADAPTIVE FILTERING
ANC because the step size must be reduced to ensure stability. Furthermore, it is not
always feasible to place the reference sensor far away from the signal source. The second
method for reducing crosstalk is to place an acoustic barrier (an oxygen mask in an
aircraft cockpit, for example) between the primary and reference sensors. However,
many applications do not allow an acoustic barrier between sensors, and a barrier may
reduce the correlation of the noise component in the primary and reference signals. The
third technique involves allowing the adaptive algorithm to update filter coefficients
during silent intervals in the speech. Unfortunately, this method depends on a reliable
speech detector that is very application dependent. This technique also fails to track the
environment changes during the speech periods.
8.5.4 Adaptive Notch Filters
In certain situations, the primary input is a broadband signal with an undesired
narrowband (sinusoidal) interference. The conventional method of eliminating
such sinusoidal interference is by using a notch filter tuned to the frequency of
the interference. To design the filter, we need to estimate the precise frequency of the
interference. A very narrow notch is usually desired in order to filter out the inter-
ference without seriously distorting the signal of interest. The advantages of the
adaptive notch filter are that it offers an infinite null, and the capability to adaptively
track the frequency of the interference. The adaptive notch filter is especially useful
when the interfering sinusoid drifts slowly in frequency. In this section, we will present
two adaptive notch filters.
The adaptive structure shown in Figure 8.7 can be applied to enhance the broadband
signal, which is corrupted by multiple narrowband components. For example, the input
signal expressed in (8.5.5) consists of a broadband signal v(n), which is music. In this
application, e(n) is the desired output that consists of enhanced broadband music
signals since the narrowband components are readjusted by W(z) so that they can
cancel correlated components in d(n). The adaptive system between the input x(n) and
the output e(n) is an adaptive notch filter, which forms several notch filters centered at
the frequency of the sinusoidal components.
A sinusoid can be used as a reference signal for canceling each component of
narrowband noise. When a sinewave is employed as the reference input, the LMS
algorithm becomes an adaptive notch filter, which removes the primary spectral com-
ponents within a narrowband centered about the reference frequency. Furthermore,
multiple harmonic disturbances can be handled if the reference signal is composed of a
number of sinusoids.
A single-frequency adaptive notch filter with two adaptive weights is illustrated in
Figure 8.10. The reference input is a cosine signal
x(n) = x
0
(n) = Acos(!
0
n): (8:5:6)
A 908 phase shifter is used to produce the quadrature reference signal
x
1
(n) = Asin(!
0
n): (8:5:7)
APPLICATIONS 377
90
Њ
phase shift
w
0
(n) w
1
(n)
+

e(n)
LMS
y(n)
x
1
(n)
d(n) x
1
(n) x
0
(n)
x
0
(n)
Figure 8.10 Single-frequency adaptive notch filter
Digital Hilbert transform filters can be employed for this purpose. Instead of using
cosine generator and a phase shifter, the recursive quadrature oscillator given in Figure
6.23 can be used to generate both sine and cosine signals simultaneously. For a reference
sinusoidal signal, two filter coefficients are needed.
The LMS algorithm employed in Figure 8.10 is summarized as follows:
y(n) = w
0
(n)x
0
(n) ÷w
1
(n)x
1
(n), (8:5:8)
e(n) = d(n) ÷y(n), (8:5:9)
w
l
(n ÷1) = w
l
(n) ÷mx
l
(n)e(n), l = 0, 1: (8:5:10)
Note that the two-weight adaptive filter W(z) shown in Figure 8.10 can be replaced with
a general L-weight adaptive FIR filter for a multiple sinusoid reference input x(n). The
reference input supplies a correlated version of the sinusoidal interference that is used to
estimate the composite sinusoidal interfering signal contained in the primary input d(n).
The single-frequency adaptive notch filter has the property of a tunable notch filter.
The center frequency of the notch filter depends on the sinusoidal reference signal,
whose frequency is equal to the frequency of the primary sinusoidal noise. Therefore the
noise at that frequency is attenuated. This adaptive notch filter provides a simple
method for tracking and eliminating sinusoidal interference.
Example 8.8: For a stationary input and sufficiently small m, the convergence
speed of the LMS algorithm is dependent on the eigenvalue spread of the input
autocorrelation matrix. For L = 2 and the reference input is given in (8.5.6),
the autocorrelation matrix can be expressed as
R = E
x
0
(n)x
0
(n) x
0
(n)x
1
(n)
x
1
(n)x
0
(n) x
1
(n)x
1
(n)
¸
= E
A
2
cos
2
(!
0
n) A
2
cos(!
0
n) sin(!
0
n)
A
2
sin(!
0
n) cos(!
0
n) A
2
sin
2
(!
0
n)
¸ ¸
=
A
2
2
0
0
A
2
2

¸
¸
¸:
378 ADAPTIVE FILTERING
This equation shows that because of the 908 phase shift, x
0
(n) is orthogonal to
x
1
(n) and the off-diagonal terms in the R matrix are 0. The eigenvalues l
1
and l
2
of the R matrix are identical and equal to A
2
=2. Therefore the system has very fast
convergence since the eigenvalue spread equals 1. The time constant of the
adaptation is approximated as
t
mse
_
1
ml
=
2
mA
2
,
which is determined by the power of the reference sinewave and the step size m.
8.5.5 Adaptive Channel Equalization
In digital communications, considerable effort has been devoted to the development of
data-transmission systems that utilize the available telephone channel bandwidth effi-
ciently. The transmission of high-speed data through a channel is limited by intersymbol
interference (ISI) caused by distortion in the transmission channel. High-speed data
transmission through channels with severe distortion can be achieved in several ways,
such as (1) by designing the transmit and receive filters so that the combination of filters
and channel results in an acceptable error from the combination of ISI and noise; and
(2) by designing an equalizer in the receiver that counteracts the channel distortion. The
second method is the most commonly used.
As illustrated in Figure 8.11, the received signal y(n) is different from the original
signal x(n) because it was distorted by the overall channel transfer function C(z), which
includes the transmit filter, the transmission medium, and the receive filter. To recover
the original signal, x(n), we need to process y(n) using the equalizer W(z), which is the
inverse of the channel's transfer function C(z) to compensate for the channel distortion.
That is, we have to design the equalizer
W(z) =
1
C(z)
, (8:5:11)
i.e., C(z)W(z) = 1 such that
^
x(n) = x(n).
In practice, the telephone channel is time varying and is unknown in the design stage
due to variations in the transmission medium. The transmit and receive filters that are

C(z) W(z)
LMS
z
−∆
d(n)
+
x(n) y(n) x(n) ˆ e(n)
Figure 8.11 Cascade of channel with an ideal adaptive channel equalizer
APPLICATIONS 379
designed based on the average channel characteristics may not adequately reduce
intersymbol interference. Thus we need an adaptive equalizer that provides precise
compensation over the time-varying channel. As illustrated in Figure 8.11, the adaptive
channel equalizer is an adaptive filter with coefficients that are adjusted using the LMS
algorithm.
As shown in Figure 8.11, an adaptive filter requires the desired signal d(n) for
computing the error signal e(n) for the LMS algorithm. In theory, the delayed version
of the transmitted signal, x(n ÷D), is the desired response for the adaptive equalizer
W(z). However, with the adaptive filter located in the receiver, the desired signal
generated by the transmitter is not available at the receiver. The desired signal may be
generated locally in the receiver using two methods. A decision-directed algorithm, in
which the equalized signal
^
x(n) is sliced to form the desired signal, is the simplest and
can be used for channels that have only a moderate amount of distortion. However, if
the error rate of the data derived by slicing is too high, the convergence may be seriously
impaired. In this case, the training of an equalizer by using that sequence is agreed on
beforehand by the transmitter and the receiver.
During the training stage, the adaptive equalizer coefficients are adjusted by trans-
mitting a short training sequence. This known transmitted sequence is also generated in
the receiver and is used as the desired signal d(n) for the LMS algorithm. A widely used
training signal consists of pseudo-random noise (will be introduced in Section 9.2) with
a broad and flat power spectrum. After the short training period, the transmitter begins
to transmit the data sequence. In order to track the possible slow time variations in the
channel, the equalizer coefficients must continue to be adjusted while receiving data. In
this data mode, the output of the equalizer,
^
x(n), is used by a decision device (slicer) to
produce binary data. Assuming the output of the decision device is correct, the binary
sequence can be used as the desired signal d(n) to generate the error signal for the LMS
algorithm.
An equalizer for a one-dimensional baseband system has real input signals and filter
coefficients. However, for a two-dimensional quadrature amplitude modulation (QAM)
system, both signals and coefficients are complex. All operations must use complex
arithmetic and the complex LMS algorithm expressed as
w(n ÷1) = w(n) ÷mx
+
(n)e(n), (8:5:12)
where * denotes complex conjugate and
x(n) = x
R
(n) ÷jx
I
(n) (8:5:13)
and
w(n) = w
R
(n) ÷jw
I
(n): (8:5:14)
In (8.5.13) and (8.5.14), the subscript R and I represent real and imaginary parts of
complex numbers. The complex output
^
x(n) is given by
^
x(n) = w
T
(n)x(n): (8:5:15)
380 ADAPTIVE FILTERING
Since all multiplications are complex, the equalizer usually requires four times as many
multiplications.
8.6 Implementation Considerations
The adaptive algorithms introduced in previous sections assume the use of in-
finite precision for the signal samples and filter coefficients. In practice, an adap-
tive algorithm is implemented on finite-precision hardware. It is important to
understand the finite-wordlength effects of adaptive algorithms in meeting design
specifications.
8.6.1 Computational Issues
The coefficient update defined in (8.2.21) requires L ÷1 multiplications and L additions
by multiplying m
+
e(n) outside the loop. Given the input vector x(n) stored in the array
x [], the error signal en, the weight vector w [], the step size mu, and the filter order L,
Equation (8.2.21) can be implemented in C as follows:
uen = mu*en; /* u*e(n) */
for(l = 0, l < L, l÷÷) /* l = 0, 1, ..., L÷1 */
{
w [l]÷= uen*x [l]; /* LMS update */
}
The architecture of most commercially available DSP chips has been optimized for
convolution operations to compute output y(n) given in (8.2.19) in L instruction cycles.
However, the weight update operations in (8.2.21) cannot take advantage of this special
architecture because each update cycle involves loading the weight value into the
accumulator, performing a multiply±add operation, and storing the result back into
memory. With typical DSP architecture, the memory transfer operations take as many
machine cycles as the multiply±add, thus resulting in a dominant computational burden
involved in the weight update.
Let T
p
denote the total processing time for each input sample. Real-time processing
requires that T
p
< T, where T is the sampling period. This means that if the algorithm is
too complicated, we need to reduce the complexity of computation to allow T
p
< T.
Because the output convolution is more important and can be implemented very
efficiently using DSP chips, skipping part of the weight update is suggested. That is,
the simplest solution is to update only a fraction of the coefficients each sample period.
The principal cost is a somewhat slower convergence rate.
In a worst-case scenario, we might update only one weight during a sample period.
The next weight would be updated for the next sample period, and so on. When the
computing power permits, a group of samples can be updated during each sample
period.
IMPLEMENTATION CONSIDERATIONS 381
8.6.2 Finite-Precision Effects
In the digital implementation of adaptive algorithms, both the signals and the internal
algorithmic quantities are carried to a certain limited precision. Therefore an adaptive
filter implementation with limited hardware precision requires special attention because
of the potential accumulation of quantization and arithmetic errors to unacceptable
levels as well as the possibility of overflow. This section analyzes finite-precision effects
in adaptive filters using fixed-point arithmetic and presents methods for confining these
effects to acceptable levels.
We assume that the input data samples are properly scaled so that their values lie
between ÷1 and 1. Each data sample and filter coefficient is represented by B bits (M
magnitude bits and one sign bit). For the addition of digital variables, the sum may
become larger than 1. This is known as overflow. As introduced in Section 3.6, the
techniques used to inhibit the probability of overflow are scaling, saturation arithmetic,
and guard bits. For adaptive filters, the feedback path makes scaling far more compli-
cated. The dynamic range of the filter output is determined by the time-varying filter
coefficients, which are unknown at the design stage.
For the adaptive FIR filter with the LMS algorithm, the scaling of the filter output
and coefficients can be achieved by scaling the `desired' signal, d(n). The scale factor a,
where 0 < a _ 1, is implemented by right-shifting the bits of the desired signal to
prevent overflow of the filter coefficients during the weight update. Reducing the
magnitude of d(n) reduces the gain demand on the filter, thereby reducing the magni-
tude of the weight values. Usually, the required value of a is not expected to be very
small. Since a only scales the desired signal, it does not affect the rate of convergence,
which depends on the reference signal x(n). An alternative method for preventing
overflow is to use the leaky LMS algorithm described in Section 8.4.2.
With rounding operations, the finite-precision LMS algorithm can be described as
follows:
y(n) = R
¸
L÷1
l=0
w
l
(n)x(n ÷l)
¸ ¸
, (8:6:1)
e(n) = R[ad(n) ÷y(n)[, (8:6:2)
w
l
(n ÷1) = R[w
l
(n) ÷mx(n ÷l)e(n)[, l = 0, 1, . . . , L ÷1, (8:6:3)
where R[x[ denotes the fixed-point rounding of the quantity x.
When the convolution sum in (8.6.1) is calculated using a multiplier with an internal
double-precision accumulator, internal roundoff noise is avoided. Therefore roundoff
error is only introduced when the product is transferred out of the accumulator and
the result is rounded to single precision. When updating weights according to (8.6.3), the
product mx(n ÷l)e(n) produces a double-precision number, which is added to the
original stored weight value, w
l
(n), then is rounded to form the updated value,
w
l
(n ÷1). Insufficient precision provided in the weight value will cause problems such
as coefficient bias or stalling of convergence, and will then be responsible for excess
error in the output of the filter.
382 ADAPTIVE FILTERING
By using the assumptions that quantization and roundoff errors are zero-mean white
noise independent of the signals and each other, that the same wordlength is used for
both signal samples and coefficients, and that m is sufficiently small, the total output
MSE is expressed as
x = x
min
÷
m
2
Ls
2
x
x
min
÷
Ls
2
e
2a
2
m
÷
1
a
2
¸

w
o

2
÷k

s
2
e
, (8:6:4)
where x
min
is the minimum MSE defined in (8.2.13), s
2
x
is the variance of the input
signal x(n), s
2
e
is as defined in (3.5.6), k is the number of rounding operations in
(8.6.1) (k = 1 if a double-precision accumulator is used), and w
o
is the optimum weight
vector.
The second term in (8.6.4) represents excess MSE due to algorithmic weight fluctu-
ation and is proportional to the step size m. For fixed-point arithmetic, the finite-
precision error given in (8.6.4) is dominated by the third term, which reflects the error
in the quantized weight vector, and is inversely proportional to the step size m. The last
term in (8.6.4) arises because of two quantization errors ± the error in the quantized
input vector and the error in the quantized filter output y(n).
Whereas the excess MSE given in the second term of (8.6.4) is proportional to m, the
power of the roundoff noise in the third term is inversely proportional to m. Although a
small value of m reduces the excess MSE, it may result in a large quantization error.
There will be an optimum step size that achieves a compromise between these competing
goals. The total error for a fixed-point implementation of the LMS algorithm is min-
imized using the optimum m
o
expressed as
m
o
=
2
÷M
2a

1
3x
min
s
2
x

: (8:6:5)
In order to stabilize the digital implementation of the LMS algorithm, we may use the
leaky LMS algorithm to reduce numeric errors accumulated in the filter coefficients. As
discussed in Section 8.4.2, the leaky LMS algorithm prevents overflow in a finite-
precision implementation by providing a compromise between minimizing the MSE
and constraining the energy of the adaptive filter impulse response.
There is still another factor to consider in the selection of step size m. As mentioned in
Section 8.2, the adaptive algorithm is aimed at minimizing the error signal, e(n). As the
weight vector converges, the error term decreases. At some point, the update term will
be rounded to 0. Since mx(n ÷l)e(n) is a gradually decreasing random quantity, fewer
and fewer values will exceed the rounding threshold level, and eventually the weight will
stop changing almost completely. The step size value m
o
given in (8.6.5) is shown to be
too small to allow the adaptive algorithm to converge completely. Thus the `optimal'
value in (8.6.5) may not be the best choice from this standpoint.
From (8.6.1)±(8.6.3), the digital filter coefficients, as well as all internal signals,
are quantized to within the least significant bit LSB = 2
÷B
. From (8.6.3), the LMS
algorithm modifies the current parameter settings by adding a correction term,
R[mx(n ÷l)e(n)[. Adaptation stops when the correction term is smaller in magnitude
than the LSB. At this point, the adaptation of the filter virtually stops. Roundoff
IMPLEMENTATION CONSIDERATIONS 383
precludes the tap weights reaching the optimum (infinite-precision) value. This phenom-
enon is known as `stalling' or `lockup'.
The condition for the lth component of the weight vector w
l
(n) not to be updated is
whenever the corresponding correction term for w
l
(n) in the update equation is smaller
in magnitude than the least significant bit of the weight:
[mx(n ÷l)e(n)[ < 2
÷M
: (8:6:6)
Suppose that this equation is first satisfied for l = 0 at time n. As the particular
input sample x(n) propagates down the tapped-delay-line, the error will further decrease
in magnitude, and thus this sample will turn off all weight adaptation beyond this
point.
To get an approximate condition for the overall algorithm to stop adapting, we can
replace [x(n ÷l)[ and [e(n)[ with their standard deviation values, s
x
and s
e
, respectively.
The condition for the adaptation to stop becomes
ms
x
s
e
< 2
÷M
: (8:6:7)
We have to select the step size m to satisfy
m _ m
min
=
2
÷M
s
x
s
e
(8:6:8)
to prevent early termination of adaptation. When adaptation is prematurely terminated
by quantization effects, the total output MSE can be decreased by increasing the step
size m.
In steady state, the optimum step size m
o
in (8.6.5) is usually smaller than the m
min
specified in (8.6.8). Therefore if the value m _ m
o
< m
min
is used, the adaptation essen-
tially stops before complete convergence of the algorithm is attained. In order to prevent
this early termination of the adaptation, some value of
m _ m
min
> m
o
(8:6:9)
is selected. In this case, the excess MSE due to misadjustment is larger than the finite-
precision error.
In conclusion, the most important design issue is to find the best value of m that
satisfies
m
min
< m <
1
Ls
2
x
: (8:6:10)
To prevent algorithm stalling due to finite-precision effects, the design must allow the
residual error to reach small non-zero values. This can be achieved by using a suffi-
ciently large number of bits, and/or using a large step size m, while still guaranteeing
convergence of the algorithm. However, this will increase excess MSE as shown in
(8.6.4).
384 ADAPTIVE FILTERING
8.7 Experiments Using the TMS320C55x
The adaptive filtering experiments we will conduct in this section are filters that adjust
their coefficients based on a given adaptive algorithm. Unlike the time-invariant filters
introduced in Chapters 5 and 6, the adaptation algorithms will optimize the filter
response based on predetermined performance criteria. The adaptive filters can be
realized as either FIR or IIR filters. As explained in Section 8.2.1, the FIR filters are
always stable and provide linear phase responses. We will conduct our experiments
using FIR filters in the subsequence sections.
As introduced in Section 8.2, there are many adaptive algorithms that can be used
for DSP applications. Among them, the LMS algorithm has been widely used in real-
time applications, such as adaptive system identification, adaptive noise cancella-
tion, channel equalization, and echo cancellation. In this section, we will use the
TMS320C55x to implement an adaptive identification system and an adaptive linear
predictor.
8.7.1 Experiment 8A ± Adaptive System Identification
The block diagram of adaptive system identification is shown in Figure 8.6. The input
sample x(n) is fed to both the unknown system and the adaptive filter. The output of the
unknown system is used by the adaptive filter as the desired signal d(n). The adaptive
algorithm minimizes the differences between the outputs of the unknown system and
adaptive filter. The filter coefficients are continuously adjusted until the error signal has
been minimized. When the adaptive filter has converged, the coefficients of the filter
describe the characteristics of the unknown system.
As shown in Figure 8.6, the system identification consists of three basic elements ± a
signal generator, an adaptive filter, and an unknown system that needs to be modeled.
The input signal x(n) should have a broad spectrum to excite all the poles and zeros of
the unknown system. Both the white noise and the chirp signal are widely used for
system identification. The signal generation algorithms will be introduced in Sections
9.1 and 9.2.
For the adaptive system identification experiment, we use the LMS algorithm
in conjunction with an FIR filter as shown in Figure 8.6. In practical applications, the
unknown system is a physical plant with both the input and output connected to
the adaptive filter. However, for experimental purposes and to better understand the
properties of adaptive algorithms, we simulate the unknown system in the same pro-
gram. The adaptive syst